Compare commits
6 Commits
3dc49683cc
...
8c8b5aa990
| Author | SHA1 | Date |
|---|---|---|
|
|
8c8b5aa990 | |
|
|
8fc2b23ccf | |
|
|
a765319bc3 | |
|
|
36ef6f6501 | |
|
|
decf31604b | |
|
|
623b1b72ef |
|
|
@ -1,555 +0,0 @@
|
|||
---
|
||||
slug: claude-virtual-filesystem-guide
|
||||
title: "Inside Claude's Sandbox: What Happens When Claude Creates a File"
|
||||
author: henry
|
||||
status: outline
|
||||
created: 2024-12-25
|
||||
updated: 2024-12-25
|
||||
content_type: debugging-story
|
||||
|
||||
primary_keyword: "claude file creation"
|
||||
secondary_keywords: ["claude sandbox", "claude mcp filesystem", "claude outputs folder", "claude virtual filesystem"]
|
||||
---
|
||||
|
||||
# Brief
|
||||
|
||||
## Strategic Context
|
||||
|
||||
### Why This Article?
|
||||
No one has documented Claude's internal sandbox filesystem structure in claude.ai. Users encounter frustration when files "disappear" or Claude creates files in the wrong location. This is Henry's first article — establishes technical credibility with an original investigation that provides real value.
|
||||
|
||||
**Availability:** Code execution and file creation requires **paid plans** (Pro, Max, Team, Enterprise). Free plan only has Artifacts.
|
||||
|
||||
### Target Reader
|
||||
- **Role:** AI-first developer using Claude Pro/Max with "Code execution and file creation" enabled
|
||||
- **Situation:** Using Claude for code generation, file creation, possibly with Filesystem MCP in Claude Desktop
|
||||
- **Pain:** Files created by Claude don't appear where expected; confusion between internal sandbox and Filesystem MCP
|
||||
- **Search query:** "claude file creation not working", "where does claude save files", "claude mcp vs sandbox"
|
||||
|
||||
### Terminology Clarification (for article)
|
||||
| Term | What it means |
|
||||
|------|---------------|
|
||||
| "Code execution and file creation" | Official Anthropic name for sandbox feature in claude.ai |
|
||||
| Sandbox / Sandboxed environment | Ubuntu container where Claude runs code |
|
||||
| Artifacts | Interactive previews (HTML, React, SVG) — separate feature from file creation |
|
||||
| Filesystem MCP | External MCP server for local file access (Claude Desktop only) |
|
||||
| "Virtual filesystem" | NOT official term, but Claude understands it in conversation — tested in practice |
|
||||
|
||||
### Success Metrics
|
||||
- Primary: Organic traffic from developers searching for Claude file issues
|
||||
- Secondary: Social shares from AI dev communities (Reddit, Twitter, Dev.to)
|
||||
|
||||
---
|
||||
|
||||
## SEO Strategy
|
||||
|
||||
### Keywords
|
||||
| Type | Keyword | Notes |
|
||||
|------|---------|-------|
|
||||
| Primary | claude file creation | High intent, problem-focused |
|
||||
| Secondary | claude sandbox environment | Technical term users encounter |
|
||||
| Secondary | claude mcp filesystem | Confusion point we address |
|
||||
| Secondary | claude virtual filesystem | Descriptive, long-tail |
|
||||
|
||||
### Search Intent
|
||||
User expects: practical explanation of where Claude stores files, how to find them, how to control file location.
|
||||
|
||||
### Competition
|
||||
- Anthropic docs exist but are high-level, don't show internal paths
|
||||
- No articles specifically about `/mnt/user-data/` structure
|
||||
- Our angle: hands-on investigation with screenshots and "try it yourself" exercises
|
||||
|
||||
### Unique Angle
|
||||
First-hand debugging story with reproducible experiments. Reader can follow along and discover the filesystem themselves.
|
||||
|
||||
---
|
||||
|
||||
## Content Requirements
|
||||
|
||||
### Core Question
|
||||
**Where do files go when Claude creates them, and how do I make Claude save files where I actually want them?**
|
||||
|
||||
### Must Cover
|
||||
1. Sandbox filesystem structure overview (key folders and their purposes)
|
||||
2. What happens when Claude creates a file (step by step)
|
||||
3. The `/mnt/user-data/outputs/` → sidebar connection
|
||||
4. Problem: Claude confusing internal sandbox vs Filesystem MCP (in Claude Desktop)
|
||||
5. Solution: how to direct Claude to the right tool
|
||||
6. **Two strategies for file workflows** (see below)
|
||||
7. "Try it yourself" experiments for readers
|
||||
8. Quick note: this requires paid plan (Pro+)
|
||||
|
||||
### Two File Workflow Strategies (new section)
|
||||
|
||||
**Strategy 1: Work in sandbox, save at end**
|
||||
- Work with files inside `/home/claude/` during conversation
|
||||
- Only move to `/mnt/user-data/outputs/` when done
|
||||
- Pros: Faster iteration, no filesystem noise, sandbox is temp anyway
|
||||
- Cons: Lose work if you forget to save, files not visible until end
|
||||
|
||||
**Strategy 2: Save to local disk immediately (via Filesystem MCP)**
|
||||
- Claude saves directly to local filesystem via MCP
|
||||
- Pros: Files persist immediately, work directly with your project files
|
||||
- Cons: Requires MCP setup in Claude Desktop, can't use in claude.ai web
|
||||
|
||||
### Must NOT Cover
|
||||
- MCP server installation guide (separate topic, just mention it exists)
|
||||
- API code execution tool (different product)
|
||||
- Artifacts deep dive (mention briefly for context on naming confusion)
|
||||
|
||||
### Note on Artifacts vs Files (sidebar box)
|
||||
Users often confuse "artifacts" and "files":
|
||||
- **Artifacts** (June 2024): Interactive previews that render in sidebar — HTML, React, SVG, code snippets
|
||||
- **Files** (September 2025): Actual downloadable documents — .docx, .xlsx, .pdf, created via sandbox
|
||||
|
||||
Artifacts had a highlight+edit feature (September 2024) where you could select code and click "Improve" or "Explain". This may have changed after the October 2025 UI update when Code Execution became default. The current interface separates Artifacts from file creation more clearly.
|
||||
|
||||
### Unique Angle
|
||||
Personal debugging story: "I spent hours confused about where my files went. Here's what I discovered."
|
||||
|
||||
### Banatie Integration
|
||||
- Type: none
|
||||
- Rationale: First Henry article, establish credibility first. No forced mentions.
|
||||
|
||||
---
|
||||
|
||||
## Structure Guidance
|
||||
|
||||
### Suggested Flow
|
||||
1. **Opening hook:** The frustration — "Claude said it created the file. But where is it?"
|
||||
2. **The investigation:** How I started exploring with `view /` commands
|
||||
3. **The map:** Key folders explained with table
|
||||
4. **The gotcha:** Sandbox vs Filesystem MCP confusion
|
||||
5. **The fix:** Specific prompts that work
|
||||
6. **Two strategies:** Sandbox-first vs Local-first workflows
|
||||
7. **Try it yourself:** Commands readers can run
|
||||
8. **Quick reference:** Cheat sheet
|
||||
|
||||
### Opening Hook
|
||||
Start with the specific frustration moment. First-person, relatable. No definitions.
|
||||
|
||||
### Closing CTA
|
||||
"Now you know where Claude keeps its files. Go explore your own sandbox — and stop losing your work."
|
||||
|
||||
---
|
||||
|
||||
## Visual & Interactive Elements
|
||||
|
||||
### Screenshots Needed
|
||||
1. Sidebar showing files in outputs folder
|
||||
2. Result of `view /mnt/user-data/` showing structure
|
||||
3. Example of Claude creating file "not in outputs" (the problem)
|
||||
|
||||
### Code Snippets for Article
|
||||
```
|
||||
view /
|
||||
view /mnt/user-data/
|
||||
view /home/claude/
|
||||
```
|
||||
|
||||
### "Try It Yourself" Exercises
|
||||
1. "Ask Claude: `view /mnt/user-data/` — what do you see?"
|
||||
2. "Ask Claude to create a test file. Check: did it appear in sidebar?"
|
||||
3. "If you have MCP configured, ask Claude to save via filesystem MCP specifically"
|
||||
|
||||
---
|
||||
|
||||
## Screenshot Flow (for Oleg to capture)
|
||||
|
||||
Create a fresh chat with Code Execution enabled. Run these in sequence:
|
||||
|
||||
```
|
||||
Step 1: "Show me the root filesystem structure with view /"
|
||||
Screenshot: The output showing available directories
|
||||
|
||||
Step 2: "Show me what's in /mnt/user-data/"
|
||||
Screenshot: uploads/, outputs/ structure
|
||||
|
||||
Step 3: "Create a simple test.txt file with 'hello world' content"
|
||||
Screenshot: Where Claude creates it (likely /home/claude/ or outputs/)
|
||||
|
||||
Step 4: "Show me /mnt/user-data/outputs/"
|
||||
Screenshot: Verify file appears (or doesn't)
|
||||
|
||||
Step 5: Check sidebar
|
||||
Screenshot: File appearing in download area
|
||||
|
||||
Step 6 (if MCP configured in Claude Desktop):
|
||||
"Use filesystem MCP to save a file to ~/Desktop/test-mcp.txt"
|
||||
Screenshot: Compare behavior — file goes to actual local disk
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## References
|
||||
|
||||
### Official Documentation
|
||||
- https://support.claude.com/en/articles/12111783-create-and-edit-files-with-claude (main reference)
|
||||
- https://docs.claude.com/en/release-notes/claude-apps (timeline of features)
|
||||
|
||||
### Research Sources
|
||||
- Personal investigation by author
|
||||
- Simon Willison's analysis: https://simonwillison.net/2025/Sep/9/claude-code-interpreter/
|
||||
|
||||
### Background Context (for author reference, not for article)
|
||||
- Artifacts launched June 2024, got highlight+edit September 2024
|
||||
- "Analysis tool" (JS-based) launched October 2024
|
||||
- Code Execution (Python/Node sandbox) replaced Analysis tool September 2025
|
||||
- October 2025: Code Execution became default for paid plans, UI changed
|
||||
- Users report highlight+edit feature may work differently now
|
||||
|
||||
### Competitor Articles
|
||||
- None directly covering this topic (unique content opportunity)
|
||||
|
||||
---
|
||||
|
||||
**Brief created:** 2024-12-25
|
||||
**Ready for:** @architect
|
||||
|
||||
---
|
||||
|
||||
# Outline
|
||||
|
||||
## Pre-Writing Notes
|
||||
|
||||
**Author:** henry
|
||||
**Voice reference:** style-guides/henry-technical.md
|
||||
**Word target:** 1200 words (range: 800-1500)
|
||||
**Content type:** debugging-story
|
||||
|
||||
Key style points from Henry's guide:
|
||||
- Opening: Start with problem/frustration, not definitions (Section 2)
|
||||
- Sections: 150-300 words, paragraphs max 4 sentences (Section 2)
|
||||
- Code ratio: 20-30% for debugging stories (Section 4)
|
||||
- Closing: Practical next step, "Go build something." (Section 2)
|
||||
- Voice: Direct, confident, first-person, "Here's the thing about..." (Section 1)
|
||||
|
||||
---
|
||||
|
||||
## Article Structure
|
||||
|
||||
### H1: Inside Claude's Sandbox: What Happens When Claude Creates a File
|
||||
*Contains primary keyword: "claude file creation"*
|
||||
|
||||
---
|
||||
|
||||
### Opening (120-150 words)
|
||||
|
||||
**Purpose:** Hook reader with professional curiosity, promise deep understanding
|
||||
|
||||
**Approach:** First-person exploration angle. Henry digs deeper because he wants to understand the system, not because he's lost.
|
||||
|
||||
**Must include:**
|
||||
- Moment of curiosity (found file easily, but wanted to understand the system)
|
||||
- Professional interest signal — "how does this actually work?"
|
||||
- Promise: "Here's what I discovered when I mapped the whole thing."
|
||||
|
||||
**Hook angle (Option A — curiosity):**
|
||||
> "Claude created the file. I found it in the sidebar in 5 seconds. But then I wondered — where is it physically? What's the filesystem structure inside? I went exploring."
|
||||
|
||||
**Hook angle (Option B — scaling up):**
|
||||
> "When you start working with Claude on real projects, you eventually hit the question: how exactly is its filesystem organized? I decided to figure it out."
|
||||
|
||||
**Transition:** "Let me show you what's actually happening under the hood."
|
||||
|
||||
---
|
||||
|
||||
### H2: The Quick Answer (80-100 words)
|
||||
|
||||
**Purpose:** Give impatient readers the core answer immediately
|
||||
|
||||
**Must cover:**
|
||||
- Claude runs in Ubuntu sandbox container
|
||||
- Key path: `/mnt/user-data/outputs/` = sidebar downloads
|
||||
- `/home/claude/` = temp workspace (disappears)
|
||||
|
||||
**Structure:**
|
||||
1. One-liner: where files actually go
|
||||
2. Why some files "disappear"
|
||||
3. Tease: "But there's more to it. Let me show you the full map."
|
||||
|
||||
**Note:** This follows Henry's "don't bury the lede" principle.
|
||||
|
||||
---
|
||||
|
||||
### H2: The Filesystem Map (200-250 words)
|
||||
|
||||
**Purpose:** Complete reference of sandbox structure
|
||||
|
||||
**Must cover:**
|
||||
- `/mnt/user-data/uploads/` — your uploaded files
|
||||
- `/mnt/user-data/outputs/` — files that appear in sidebar
|
||||
- `/home/claude/` — temp working directory
|
||||
- `/mnt/skills/` — Claude's built-in capabilities
|
||||
- Note: `/mnt/project/` for Projects feature
|
||||
|
||||
**Structure:**
|
||||
1. Brief intro: "Here's the full structure I mapped out."
|
||||
2. Table with 4-5 key directories
|
||||
3. Key insight: only `/mnt/user-data/outputs/` = downloadable
|
||||
|
||||
**Table format:**
|
||||
| Path | What's there | Persists? |
|
||||
|------|--------------|-----------||
|
||||
| `/mnt/user-data/uploads/` | Your uploaded files | Session |
|
||||
| `/mnt/user-data/outputs/` | Files for download | Session |
|
||||
| `/home/claude/` | Claude's workspace | ❌ No |
|
||||
| `/mnt/skills/` | Built-in capabilities | Read-only |
|
||||
|
||||
**Code example:**
|
||||
```
|
||||
view /mnt/user-data/
|
||||
```
|
||||
Shows: uploads/, outputs/ structure
|
||||
|
||||
**Transition:** "So why do files sometimes not appear in your sidebar?"
|
||||
|
||||
---
|
||||
|
||||
### H2: The Problem: Where Did My File Go? (150-180 words)
|
||||
|
||||
**Purpose:** Explain the common frustration point
|
||||
|
||||
**Must cover:**
|
||||
- Claude sometimes creates in `/home/claude/` (temp, not visible)
|
||||
- Files there won't appear in sidebar
|
||||
- Claude "forgets" to move to outputs
|
||||
- Second issue: Filesystem MCP confusion (Claude Desktop)
|
||||
|
||||
**Structure:**
|
||||
1. The problem: Claude creates file in wrong place
|
||||
2. Why it happens: `/home/claude/` is default working dir
|
||||
3. Extra confusion: MCP vs sandbox (brief mention)
|
||||
|
||||
**Key insight:**
|
||||
> "Claude's sandbox resets between conversations. If a file is in `/home/claude/` and you close the chat — it's gone."
|
||||
|
||||
**⚠️ TODO:** Verify this claim with testing. Need to confirm sandbox reset behavior.
|
||||
|
||||
**Transition:** "Here's how to make sure your files end up where you can actually get them."
|
||||
|
||||
---
|
||||
|
||||
### H2: The Fix: How to Direct Claude (150-200 words)
|
||||
|
||||
**Purpose:** Give actionable solution
|
||||
|
||||
**Must cover:**
|
||||
- Explicit instruction: "save to /mnt/user-data/outputs/"
|
||||
- Example prompts that work
|
||||
- For MCP users: specify "use filesystem MCP" vs sandbox
|
||||
|
||||
**Structure:**
|
||||
1. What to say to Claude (prompt examples)
|
||||
2. For MCP users: disambiguation
|
||||
|
||||
**Prompts that work:**
|
||||
```
|
||||
"Create the file and save it to /mnt/user-data/outputs/"
|
||||
```
|
||||
|
||||
```
|
||||
"Copy this file to /mnt/user-data/outputs/"
|
||||
```
|
||||
|
||||
```
|
||||
"Use filesystem MCP to save to ~/Projects/myapp/image.png"
|
||||
```
|
||||
|
||||
**Transition:** "Now, which approach should you actually use?"
|
||||
|
||||
---
|
||||
|
||||
### H2: Two Strategies for File Workflows (200-250 words)
|
||||
|
||||
**Purpose:** Help reader choose their approach based on workflow type
|
||||
|
||||
**Must cover:**
|
||||
- Strategy 1: Sandbox for iterative work
|
||||
- Strategy 2: MCP for automation
|
||||
- When to use each (clear criteria)
|
||||
|
||||
**Structure:**
|
||||
|
||||
**Strategy 1: Sandbox-first (iterative editing)**
|
||||
- Work in `/home/claude/` during conversation
|
||||
- Use `str_replace` tool for line-by-line edits
|
||||
- Copy to outputs when done
|
||||
- Pros: faster iteration, built-in editing tools, no filesystem noise
|
||||
- Cons: files not visible in sidebar until you copy them out
|
||||
- Best for: iterative work on a single file, multiple rounds of edits, refactoring
|
||||
|
||||
**Strategy 2: MCP-first (automation)**
|
||||
- Claude saves directly to local filesystem via MCP
|
||||
- Pros: files persist immediately in your project, no extra step
|
||||
- Cons: no `str_replace` tool, requires MCP setup, Claude Desktop only
|
||||
- Best for: generating multiple files at once, scaffolding, automated workflows
|
||||
|
||||
**Key difference to highlight:**
|
||||
> Sandbox has `str_replace` for precise line-by-line editing. MCP doesn't. Choose based on whether you need iteration or automation.
|
||||
|
||||
**One-liner summary:**
|
||||
> "Sandbox for iteration. MCP for automation."
|
||||
|
||||
---
|
||||
|
||||
### H2: Try It Yourself (100-130 words)
|
||||
|
||||
**Purpose:** Reader engagement, verification
|
||||
|
||||
**Must cover:**
|
||||
- 3 quick commands to explore their own sandbox
|
||||
- What to look for
|
||||
|
||||
**Exercises:**
|
||||
1. `"Show me view /mnt/user-data/"` — see your structure
|
||||
2. `"Create test.txt with 'hello' and show me where it went"` — test file creation
|
||||
3. `"List contents of /home/claude/"` — see temp workspace
|
||||
|
||||
**Note:** Remind that this requires Pro+ plan.
|
||||
|
||||
---
|
||||
|
||||
### H2: Project Instructions for File Handling (150-180 words)
|
||||
|
||||
**Purpose:** Give readers ready-to-use instructions they can add to Claude Projects
|
||||
|
||||
**Must cover:**
|
||||
- Example instruction for sandbox-first workflow
|
||||
- Example instruction for MCP-first workflow
|
||||
- How to specify which tool to use
|
||||
|
||||
**Example 1 — Sandbox-first (for iterative work):**
|
||||
```
|
||||
File handling:
|
||||
- Work with files in /home/claude/ during conversation
|
||||
- Use str_replace for edits
|
||||
- Copy final versions to /mnt/user-data/outputs/ before finishing
|
||||
```
|
||||
|
||||
**Example 2 — MCP-first (for automation):**
|
||||
```
|
||||
File handling:
|
||||
- Use filesystem MCP to save files directly to project directory
|
||||
- Do not use sandbox for file operations
|
||||
- Save to: ~/Projects/[project-name]/
|
||||
```
|
||||
|
||||
**Example 3 — Hybrid (explicit routing):**
|
||||
```
|
||||
File handling:
|
||||
- For iterative editing: use sandbox + str_replace, copy to outputs when done
|
||||
- For generating new files: use filesystem MCP to save directly to ~/Projects/
|
||||
- Always confirm which method before creating files
|
||||
```
|
||||
|
||||
**Note:** These go in Project Instructions or system prompt.
|
||||
|
||||
---
|
||||
|
||||
### Callout Box: Artifacts ≠ Files (60-80 words)
|
||||
|
||||
**Purpose:** Address common terminology confusion
|
||||
|
||||
**Placement:** As sidebar/callout, possibly after "The Fix" or near end
|
||||
|
||||
**Must clarify:**
|
||||
- Artifacts: interactive previews (HTML, React, SVG) — render in sidebar
|
||||
- Files: actual downloadable documents (.docx, .xlsx, .pdf)
|
||||
- Different features, different behavior
|
||||
|
||||
---
|
||||
|
||||
### Closing (80-100 words)
|
||||
|
||||
**Purpose:** Wrap up with practical takeaway
|
||||
|
||||
**Approach:** Henry-style direct ending. No fluff.
|
||||
|
||||
**Must include:**
|
||||
- One-sentence summary of key insight
|
||||
- Clear CTA (explore your sandbox)
|
||||
- Sign-off phrase
|
||||
|
||||
**Draft closing:**
|
||||
> "That's it. Claude's sandbox isn't magic — it's just Ubuntu with a specific folder structure. Know the paths, and you'll never lose a file again."
|
||||
>
|
||||
> "Now go explore your own sandbox. And maybe save that important file before you close the chat."
|
||||
|
||||
---
|
||||
|
||||
## Word Count Breakdown
|
||||
|
||||
| Section | Words |
|
||||
|---------|-------|
|
||||
| Opening | 130 |
|
||||
| The Quick Answer | 90 |
|
||||
| The Filesystem Map | 220 |
|
||||
| The Problem | 160 |
|
||||
| The Fix | 180 |
|
||||
| Two Strategies | 230 |
|
||||
| Try It Yourself | 110 |
|
||||
| Project Instructions | 160 |
|
||||
| Callout: Artifacts ≠ Files | 70 |
|
||||
| Closing | 90 |
|
||||
| **Total** | **~1440** |
|
||||
|
||||
*Target: 1200-1500 (debugging story range) ✓*
|
||||
|
||||
---
|
||||
|
||||
## Code Examples Plan
|
||||
|
||||
| Section | Type | Purpose | Lines |
|
||||
|---------|------|---------|-------|
|
||||
| Filesystem Map | Command | Show structure | 1 |
|
||||
| Filesystem Map | Output | Example result | 4-5 |
|
||||
| The Fix | Prompt | Working instruction | 1 |
|
||||
| The Fix | Prompt | MCP instruction | 1 |
|
||||
| Try It Yourself | Commands | Reader exercises | 3 |
|
||||
|
||||
*Code ratio: ~15-20% (appropriate for debugging story)*
|
||||
|
||||
---
|
||||
|
||||
## Visual Elements Plan
|
||||
|
||||
| Element | Section | Description |
|
||||
|---------|---------|-------------|
|
||||
| Screenshot 1 | Filesystem Map | Output of `view /mnt/user-data/` |
|
||||
| Screenshot 2 | The Problem | File in sidebar (successful) |
|
||||
| Screenshot 3 | Try It Yourself | Optional: annotated sandbox structure |
|
||||
| Table | Filesystem Map | Directory reference |
|
||||
| Callout box | After The Fix | Artifacts vs Files clarification |
|
||||
|
||||
---
|
||||
|
||||
## SEO Notes
|
||||
|
||||
- [x] H1 contains: "Claude" + "File" (variant of primary keyword)
|
||||
- [x] H2s with keywords: "Filesystem", "File Workflows"
|
||||
- [x] First 100 words include: "Claude", "file", "created" (primary keyword area)
|
||||
- [ ] Meta description: @writer to draft — focus on "where Claude saves files"
|
||||
|
||||
---
|
||||
|
||||
## Quality Gates for @writer
|
||||
|
||||
Before submitting draft:
|
||||
- [ ] Opening starts with curiosity/professional interest, not frustration or confusion
|
||||
- [ ] "Here's the thing..." or similar Henry phrase used
|
||||
- [ ] All "Must include" items covered
|
||||
- [ ] Word counts within range per section
|
||||
- [ ] Table in Filesystem Map section present
|
||||
- [ ] Code examples complete and accurate
|
||||
- [ ] Project Instructions section has 3 ready-to-use examples
|
||||
- [ ] Callout box for Artifacts/Files distinction included
|
||||
- [ ] Closing has practical CTA, no fluff
|
||||
- [ ] First-person voice throughout
|
||||
- [ ] No forbidden phrases (see Henry guide)
|
||||
- [ ] ⚠️ Sandbox reset claim verified before publishing
|
||||
|
||||
---
|
||||
|
||||
**Outline created:** 2024-12-25
|
||||
**Ready for:** @writer
|
||||
|
After Width: | Height: | Size: 93 KiB |
|
After Width: | Height: | Size: 94 KiB |
|
After Width: | Height: | Size: 96 KiB |
|
After Width: | Height: | Size: 310 KiB |
|
After Width: | Height: | Size: 96 KiB |
|
After Width: | Height: | Size: 175 KiB |
|
After Width: | Height: | Size: 247 KiB |
|
After Width: | Height: | Size: 274 KiB |
|
After Width: | Height: | Size: 62 KiB |
|
|
@ -0,0 +1,487 @@
|
|||
<img src="https://r2cdn.perplexity.ai/pplx-full-logo-primary-dark%402x.png" style="height:64px;margin-right:32px"/>
|
||||
|
||||
# что с A2A протоколом на январь 2026? что то не слышно про его широкое использование
|
||||
|
||||
Вы правы в своих наблюдениях. По состоянию на январь 2026 года протокол A2A (Agent-to-Agent) от Google действительно не получил широкого распространения, несмотря на громкий запуск и серьезную корпоративную поддержку.
|
||||
|
||||
## Текущая ситуация с A2A
|
||||
|
||||
### Что произошло с протоколом
|
||||
|
||||
A2A был анонсирован Google в апреле 2025 года с большими амбициями — стать универсальным стандартом для взаимодействия AI-агентов[^1_1]. В июне 2025 года протокол был передан в Linux Foundation, а в июле вышла версия 0.3 с поддержкой gRPC и улучшенной безопасностью[^1_2][^1_3]. К этому моменту Google заявляла о поддержке более 150 организаций, включая Adobe, ServiceNow, S\&P Global и Twilio[^1_2][^1_3].
|
||||
|
||||
Однако к сентябрю 2025 года стало очевидно, что протокол теряет импульс. Разработка значительно замедлилась, а в качестве де-факто стандарта закрепился Model Context Protocol (MCP) от Anthropic[^1_3].
|
||||
|
||||
## Почему A2A не взлетел
|
||||
|
||||
### 1. **Чрезмерная сложность для разработчиков**
|
||||
|
||||
A2A пытался решить все проблемы сразу: обнаружение агентов, обмен возможностями, долгосрочные задачи, обработку ошибок, даже обновления UI[^1_4]. CTO HubSpot Dharmesh Shah назвал протокол "too much" — слишком сложным для практического использования[^1_4].
|
||||
|
||||
Для индивидуального разработчика, желающего создать простую интеграцию за выходные, A2A был практически недоступен[^1_3]. Требовалось понимание:
|
||||
|
||||
- Сложных паттернов оркестрации агентов
|
||||
- Реализации security cards
|
||||
- Многоагентных коммуникационных протоколов
|
||||
- Корпоративных стратегий развертывания
|
||||
|
||||
|
||||
### 2. **Неправильная стратегия выхода на рынок**
|
||||
|
||||
Google выбрала подход "сверху вниз" (top-down), ориентируясь на корпоративный сегмент, в то время как MCP использовал подход "снизу вверх" (bottom-up), завоевывая разработчиков[^1_3].
|
||||
|
||||
MCP был интегрирован с Claude с первого дня, что позволяло любому разработчику мгновенно протестировать свои серверы с production-ready AI-ассистентом[^1_3]. A2A же требовал построения новой инфраструктуры без немедленной пользы.
|
||||
|
||||
### 3. **Отсутствие ключевых игроков**
|
||||
|
||||
Критически важно: OpenAI, Anthropic и Hugging Face не присоединились к инициативе[^1_4]. Без поддержки крупных провайдеров моделей протокол не мог стать универсальным стандартом.
|
||||
|
||||
### 4. **Проблемы с инфраструктурой**
|
||||
|
||||
A2A не предоставил:
|
||||
|
||||
- Единого каталога агентов
|
||||
- Масштабируемой системы обнаружения (discovery)
|
||||
- Универсального маркетплейса[^1_4]
|
||||
|
||||
Модель обнаружения агентов изначально не была спроектирована для работы в масштабе[^1_4].
|
||||
|
||||
## MCP выиграл битву за разработчиков
|
||||
|
||||
### Почему MCP победил
|
||||
|
||||
**Статистика роста MCP (по данным на конец 2025):**
|
||||
|
||||
- 5,500+ серверов в реестрах[^1_5]
|
||||
- 180,000+ месячных поисковых запросов по топ MCP-серверам[^1_5]
|
||||
- 80% самых популярных MCP-серверов предлагают remote deployment[^1_5]
|
||||
- 4x рост remote-серверов с мая 2025[^1_5]
|
||||
|
||||
**Прогнозы Gartner:** к 2026 году 75% вендоров API-шлюзов и 50% iPaaS-провайдеров будут иметь функции MCP[^1_6].
|
||||
|
||||
### Ключевые преимущества MCP
|
||||
|
||||
1. **Философия developer-first**: простота и немедленная польза[^1_3]
|
||||
2. **Прагматичная простота**: решает реальные потребности (чтение файлов, API-вызовы, скрипты)[^1_3]
|
||||
3. **Эволюция на основе обратной связи**: быстрая адаптация к запросам сообщества[^1_3]
|
||||
4. **Интеграция с популярными инструментами**: Claude, VS Code, AWS[^1_7]
|
||||
|
||||
## Реальные примеры использования A2A
|
||||
|
||||
Несмотря на проблемы, есть несколько подтвержденных кейсов:
|
||||
|
||||
1. **Tyson Foods и Gordon Food Service** — оптимизация цепочек поставок через обмен данными о продуктах[^1_2][^1_3]
|
||||
2. **Databricks** — пример развертывания A2A-серверов как приложений (ноябрь 2025)[^1_8]
|
||||
3. **AG2 (бывший AutoGen)** — нативная поддержка A2A в фреймворке с 3,700+ звездами на GitHub[^1_9]
|
||||
4. **ServiceNow** — аутентификация для Google A2A в релизе декабря 2025[^1_10]
|
||||
|
||||
Однако в Reddit-сообществах консультанты жалуются, что не могут найти реальные production use cases с измеримым ROI[^1_11][^1_12].
|
||||
|
||||
## Технические проблемы
|
||||
|
||||
### Безопасность и управление
|
||||
|
||||
- Слабый контроль времени жизни токенов[^1_13]
|
||||
- Проблемы с compliance при обмене чувствительными данными между агентами[^1_14]
|
||||
- Сложности с observability в асинхронных цепочках агентов[^1_14]
|
||||
- Schema drift при эволюции агентов[^1_14]
|
||||
|
||||
|
||||
### Архитектурные ограничения
|
||||
|
||||
- HTTP request/response не идеален для сложных персистентных взаимодействий агентов[^1_15]
|
||||
- Квадратичная сложность подключений (O(n²)) при росте числа агентов[^1_16]
|
||||
- Жесткая связанность через прямые HTTP-соединения[^1_16]
|
||||
|
||||
|
||||
## Дополнительная перспектива
|
||||
|
||||
Интересно, что даже Google Cloud начала добавлять совместимость с MCP в свои AI-сервисы — это молчаливое признание выбора сообщества разработчиков[^1_3].
|
||||
|
||||
**Ключевой урок:** "Опыт разработчика важнее функциональности" (Developer experience trumps features)[^1_3]. MCP победил не потому, что был архитектурно лучше, а потому что разработчики действительно хотели его использовать.
|
||||
|
||||
## Будущее A2A
|
||||
|
||||
A2A технически не мёртв — протокол существует, спецификация доступна, есть поддержка Linux Foundation. Но:
|
||||
|
||||
- Разработка значительно замедлилась[^1_3]
|
||||
- Большинство AI-экосистемы консолидировалась вокруг MCP[^1_3]
|
||||
- Видение взаимодействующих агентов теперь реализуется на базе MCP, а не A2A[^1_3]
|
||||
|
||||
A2A, скорее всего, останется нишевым корпоративным решением для специфических сценариев многоагентной оркестрации в крупных организациях, которым нужны его enterprise-возможности.
|
||||
|
||||
***
|
||||
|
||||
**Вывод:** Ваша интуиция была правильной. Несмотря на мощную поддержку Google, 150+ партнерских организаций и солидную техническую архитектуру, A2A не смог завоевать сердца разработчиков. Протокол оказался over-engineered для большинства реальных задач, недоступным для индивидуальных разработчиков и опоздал на рынок, где MCP уже захватил критическую массу пользователей.
|
||||
<span style="display:none">[^1_17][^1_18][^1_19][^1_20][^1_21][^1_22][^1_23][^1_24][^1_25][^1_26][^1_27][^1_28][^1_29][^1_30][^1_31][^1_32][^1_33][^1_34][^1_35][^1_36][^1_37][^1_38][^1_39][^1_40][^1_41][^1_42][^1_43][^1_44][^1_45][^1_46][^1_47][^1_48][^1_49][^1_50][^1_51][^1_52][^1_53][^1_54][^1_55][^1_56][^1_57][^1_58][^1_59][^1_60][^1_61][^1_62][^1_63][^1_64][^1_65][^1_66][^1_67][^1_68][^1_69][^1_70][^1_71][^1_72][^1_73][^1_74][^1_75][^1_76][^1_77][^1_78][^1_79][^1_80][^1_81][^1_82][^1_83][^1_84][^1_85][^1_86][^1_87][^1_88][^1_89][^1_90][^1_91][^1_92][^1_93][^1_94][^1_95]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^1_1]: https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/
|
||||
|
||||
[^1_2]: https://cloud.google.com/blog/products/ai-machine-learning/agent2agent-protocol-is-getting-an-upgrade
|
||||
|
||||
[^1_3]: https://blog.fka.dev/blog/2025-09-11-what-happened-to-googles-a2a/
|
||||
|
||||
[^1_4]: https://www.mindset.ai/blogs/in-the-loop-ep12-what-is-a2a-protocol
|
||||
|
||||
[^1_5]: https://mcpmanager.ai/blog/mcp-adoption-statistics/
|
||||
|
||||
[^1_6]: https://www.k2view.com/blog/mcp-gartner/
|
||||
|
||||
[^1_7]: https://www.iamdave.ai/blog/top-10-model-context-protocol-use-cases-complete-guide-for-2025/
|
||||
|
||||
[^1_8]: https://community.databricks.com/t5/technical-blog/how-to-deploy-agent-to-agent-a2a-protocol-on-databricks-apps-gt/ba-p/134213
|
||||
|
||||
[^1_9]: https://discuss.google.dev/t/introducing-native-a2a-protocol-support-in-ag2-building-interoperable-multi-agent-systems-at-scale/286168
|
||||
|
||||
[^1_10]: https://www.servicenow.com/community/now-assist-articles/authentication-for-google-a2a-servicenow-as-secondary-agent/ta-p/3446091
|
||||
|
||||
[^1_11]: https://www.reddit.com/r/mcp/comments/1kg66yc/consultant_here_any_actual_a2a_use_cases_running/
|
||||
|
||||
[^1_12]: https://www.reddit.com/r/AgentToAgent/comments/1kg63rc/consultant_here_any_actual_a2a_use_cases_running/
|
||||
|
||||
[^1_13]: https://arxiv.org/html/2505.12490v3
|
||||
|
||||
[^1_14]: https://www.infobip.com/glossary/a2a-agent-to-agent
|
||||
|
||||
[^1_15]: https://www.tmasolutions.com/insights/agent2agent-a2a-protocol-opening-a-new-era-for-autonomous-ai-agents-interaction
|
||||
|
||||
[^1_16]: https://www.hivemq.com/blog/a2a-enterprise-scale-agentic-ai-collaboration-part-1/
|
||||
|
||||
[^1_17]: https://onereach.ai/blog/what-is-a2a-agent-to-agent-protocol/
|
||||
|
||||
[^1_18]: https://www.apono.io/blog/what-is-agent2agent-a2a-protocol-and-how-to-adopt-it/
|
||||
|
||||
[^1_19]: https://a2a-protocol.org/latest/roadmap/
|
||||
|
||||
[^1_20]: https://www.adopt.ai/blog/mcp-vs-a2a-in-practice
|
||||
|
||||
[^1_21]: https://datatracker.ietf.org/doc/draft-yang-nmrg-a2a-nm/
|
||||
|
||||
[^1_22]: https://research.aimultiple.com/agent2agent/
|
||||
|
||||
[^1_23]: https://www.everestgrp.com/uncategorized/the-rise-of-agent-protocols-exploring-mcp-a2a-and-acp-blog.html
|
||||
|
||||
[^1_24]: https://a2a-protocol.org/latest/
|
||||
|
||||
[^1_25]: https://www.a2aprotocol.net/en
|
||||
|
||||
[^1_26]: https://www.gravitee.io/blog/googles-agent-to-agent-a2a-and-anthropics-model-context-protocol-mcp
|
||||
|
||||
[^1_27]: https://www.linkedin.com/pulse/introduction-agent-a2a-protocol-aneshka-goyal-ehvsc
|
||||
|
||||
[^1_28]: https://www.youtube.com/watch?v=4fNUpRKVJ7A
|
||||
|
||||
[^1_29]: https://a2a-protocol.org/latest/specification/
|
||||
|
||||
[^1_30]: https://www.elastic.co/search-labs/blog/a2a-protocol-mcp-llm-agent-workflow-elasticsearch
|
||||
|
||||
[^1_31]: https://github.com/a2aproject/A2A
|
||||
|
||||
[^1_32]: https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/05/07/empowering-multi-agent-apps-with-the-open-agent2agent-a2a-protocol/
|
||||
|
||||
[^1_33]: https://www.a2aprotocol.org/en
|
||||
|
||||
[^1_34]: https://www.devshorts.in/p/agent2agent-a2a-protocol-explained
|
||||
|
||||
[^1_35]: https://www.linkedin.com/pulse/agent2agent-protocol-analysis-enterprise-ai-game-changer-treiber-z0h8f
|
||||
|
||||
[^1_36]: https://www.blott.com/blog/post/how-the-agent2agent-protocol-a2a-actually-works-a-technical-breakdown
|
||||
|
||||
[^1_37]: https://thedataexchange.media/a2a-protocol/
|
||||
|
||||
[^1_38]: https://dev.to/czmilo/2025-complete-guide-agent2agent-a2a-protocol-the-new-standard-for-ai-agent-collaboration-1pph
|
||||
|
||||
[^1_39]: https://www.leanware.co/insights/a2a-vs-acp-key-differences-developer-insights-real-world-use-cases
|
||||
|
||||
[^1_40]: https://www.cybage.com/blog/mastering-google-s-a2a-protocol-the-complete-guide-to-agent-to-agent-communication
|
||||
|
||||
[^1_41]: https://strandsagents.com/latest/documentation/docs/user-guide/concepts/multi-agent/agent-to-agent/
|
||||
|
||||
[^1_42]: https://a2a.how/applications
|
||||
|
||||
[^1_43]: https://docs.cloud.google.com/run/docs/deploy-a2a-agents
|
||||
|
||||
[^1_44]: https://trickle.so/blog/how-google-a2a-protocol-actually-works
|
||||
|
||||
[^1_45]: https://github.com/a2aproject/a2a-samples
|
||||
|
||||
[^1_46]: https://www.ibm.com/think/tutorials/use-a2a-protocol-for-ai-agent-communication
|
||||
|
||||
[^1_47]: https://blog.fka.dev/blog/2025-04-15-why-googles-a2a-protocol-doesnt-make-sense/
|
||||
|
||||
[^1_48]: https://www.clarifai.com/blog/mcp-vs-a2a-clearly-explained
|
||||
|
||||
[^1_49]: https://www.kai-waehner.de/blog/2025/05/26/agentic-ai-with-the-agent2agent-protocol-a2a-and-mcp-using-apache-kafka-as-event-broker/
|
||||
|
||||
[^1_50]: https://blog.logto.io/a2a-mcp
|
||||
|
||||
[^1_51]: https://onereach.ai/blog/power-of-multi-agent-ai-open-protocols/
|
||||
|
||||
[^1_52]: https://zbrain.ai/understanding-the-a2a-protocol/
|
||||
|
||||
[^1_53]: https://www.firecrawl.dev/blog/mcp-vs-a2a-protocols
|
||||
|
||||
[^1_54]: https://arxiv.org/abs/2505.03864
|
||||
|
||||
[^1_55]: https://www.descope.com/blog/post/mcp-vs-a2a
|
||||
|
||||
[^1_56]: https://www.legionsecurity.ai/blog-posts/security-analysis-ai-protocols
|
||||
|
||||
[^1_57]: https://www.koyeb.com/blog/a2a-and-mcp-start-of-the-ai-agent-protocol-wars
|
||||
|
||||
[^1_58]: https://www.youtube.com/watch?v=c-jbiw3QM4o
|
||||
|
||||
[^1_59]: https://oxylabs.io/blog/mcp-vs-a2a
|
||||
|
||||
[^1_60]: https://kodekloud.com/blog/a2a-protocol/
|
||||
|
||||
[^1_61]: https://www.linkedin.com/posts/suchisubhra-sinha-07b29654_announcing-the-agent2agent-protocol-a2a-activity-7319780945853652992-Dn8T
|
||||
|
||||
[^1_62]: https://www.reddit.com/r/LLMDevs/comments/1jvij21/awesome_a2a_a_curated_list_of_agent2agent/
|
||||
|
||||
[^1_63]: https://pypi.org/project/python-a2a/
|
||||
|
||||
[^1_64]: https://a2a-protocol.org/dev/community/
|
||||
|
||||
[^1_65]: https://a2a-protocol.org/latest/community/
|
||||
|
||||
[^1_66]: https://github.com/a2aproject
|
||||
|
||||
[^1_67]: https://aws.plainenglish.io/the-a2a-protocol-reality-check-what-google-isnt-telling-you-c273f5e39911
|
||||
|
||||
[^1_68]: https://github.com/a2aproject/a2a-python
|
||||
|
||||
[^1_69]: https://www.linkedin.com/posts/ai-security_im-not-a-fortune-teller-but-heres-my-educated-activity-7363574219336474629-0Gls
|
||||
|
||||
[^1_70]: https://www.youtube.com/watch?v=mFkw3p5qSuA
|
||||
|
||||
[^1_71]: https://github.com/a2aproject/a2a-js
|
||||
|
||||
[^1_72]: https://www.cohorte.co/blog/comparing-anthropics-model-context-protocol-mcp-vs-googles-agent-to-agent-a2a-for-ai-agents-in-business-automation
|
||||
|
||||
[^1_73]: https://www.anthropic.com/news/model-context-protocol
|
||||
|
||||
[^1_74]: https://www.thoughtworks.com/insights/blog/generative-ai/model-context-protocol-beneath-hype
|
||||
|
||||
[^1_75]: https://www.youtube.com/watch?v=L6OhsaqOgW0
|
||||
|
||||
[^1_76]: https://evergreen.insightglobal.com/the-new-model-context-protocol-for-ai-agents/
|
||||
|
||||
[^1_77]: https://www.linkedin.com/posts/leadgenmanthan_anthropics-mcp-vs-googles-a2a-vs-ibms-activity-7327551934230364161-slzH
|
||||
|
||||
[^1_78]: https://www.thoughtworks.com/insights/blog/generative-ai/model-context-protocol-mcp-impact-2025
|
||||
|
||||
[^1_79]: https://modelcontextprotocol.io/development/roadmap
|
||||
|
||||
[^1_80]: https://www.youtube.com/watch?v=wrCF8MoXC_I
|
||||
|
||||
[^1_81]: https://www.iansresearch.com/what-we-do/events/symposiums/details/2026/02/19/2026-symposium/february-19-mcp-risks-and-opportunities-in-an-ai-world
|
||||
|
||||
[^1_82]: https://modelcontextprotocol.io/specification/2025-11-25
|
||||
|
||||
[^1_83]: https://guptadeepak.com/a-comparative-analysis-of-anthropics-model-context-protocol-and-googles-agent-to-agent-protocol/
|
||||
|
||||
[^1_84]: https://www.cdata.com/blog/2026-year-enterprise-ready-mcp-adoption
|
||||
|
||||
[^1_85]: https://a16z.com/a-deep-dive-into-mcp-and-the-future-of-ai-tooling/
|
||||
|
||||
[^1_86]: https://www.strategysoftware.com/blog/model-context-protocol-mcp-for-enterprise-ai-integration
|
||||
|
||||
[^1_87]: https://devops.com/how-model-context-protocol-mcp-is-fueling-the-next-era-of-developer-productivity/
|
||||
|
||||
[^1_88]: https://www.linkedin.com/posts/samcharrington_ai-agents-a2a-activity-7315791395498188801-EX_W
|
||||
|
||||
[^1_89]: https://www.altexsoft.com/blog/a2a-protocol-explained/
|
||||
|
||||
[^1_90]: https://pub.towardsai.net/we-spent-47-000-running-ai-agents-in-production-heres-what-nobody-tells-you-about-a2a-and-mcp-5f845848de33
|
||||
|
||||
[^1_91]: https://dev.to/sreeni5018/understanding-googles-a2a-protocol-the-future-of-agent-communication-part-i-334p
|
||||
|
||||
[^1_92]: https://www.finextra.com/blogposting/30530/agent-to-agent-protocol-when-software-learns-to-talk-to-itself-politely
|
||||
|
||||
[^1_93]: https://www.leanware.co/insights/a2a-vs-mcp-protocol-comparison
|
||||
|
||||
[^1_94]: https://dev.to/czmilo/impact-analysis-google-donating-a2a-protocol-to-linux-foundation-3efc
|
||||
|
||||
[^1_95]: https://dev.to/blurbify/why-you-should-care-about-the-agent2agent-a2a-protocol-as-a-developer-13n4
|
||||
|
||||
|
||||
---
|
||||
|
||||
# A2A и MCP - это чтоли конкурирующие протоколы?
|
||||
|
||||
По замыслу — **скорее дополняющие**, по факту рынка — выглядят как конкуренты за “стандарт”, но решают разные уровни задачи.
|
||||
|
||||
## Что делает MCP
|
||||
|
||||
- MCP стандартизирует, **как один агент/LLM ходит к тулзам и данным**: файлы, БД, внешние HTTP API, скрипты, вьюхи и т.д.[^2_1][^2_2]
|
||||
- Это вертикальная интеграция: “модель ↔ инструменты/контекст”.[^2_3]
|
||||
- Типичный сценарий: Claude/другой LLM через MCP обращается к твоим сервисам и источникам данных, но сам по себе не регламентирует общение между разными агентами.[^2_1][^2_2]
|
||||
|
||||
|
||||
## Что делает A2A
|
||||
|
||||
- A2A стандартизирует, **как агенты общаются друг с другом**: описывают свои способности (agent cards), ставят задачи друг другу, обмениваются артефактами.[^2_1][^2_2][^2_3]
|
||||
- Это горизонтальный слой: “агент ↔ агент”, оркестрация и многоагентные воркфлоу.[^2_4][^2_3]
|
||||
- Типичный сценарий: несколько специализированных агентов из разных систем координируют долгий процесс (R\&D, supply chain, онбординг) через A2A.[^2_5][^2_6]
|
||||
|
||||
|
||||
## Официальная позиция и практика
|
||||
|
||||
- И Google, и Anthropic публично говорят: протоколы **должны работать вместе, а не вместо** друг друга.[^2_7][^2_2][^2_8]
|
||||
- Часто рисуют такую картинку:
|
||||
- MCP даёт каждому агенту доступ к своим инструментам.
|
||||
- A2A позволяет этим “прокачанным” агентам координироваться между собой.[^2_1][^2_5][^2_9]
|
||||
|
||||
Пример “идеальной” архитектуры из гайдов:
|
||||
|
||||
- Внутри каждого сервиса агент использует MCP-серверы для доступа к БД, CRM, платежам.
|
||||
- Между сервисами агенты общаются по A2A, передавая задачи и результаты.[^2_5][^2_10]
|
||||
|
||||
|
||||
## Почему всё же говорят о конкуренции
|
||||
|
||||
- На практике компании выбирают, **во что инвестировать в первую очередь**, и MCP уже стал де‑факто стандартом для dev‑экосистемы, IDE и ассистентов.[^2_11][^2_10]
|
||||
- Из‑за этого A2A воспринимается как “конкурент”, хотя по спецификации он решает другой слой задачи.[^2_7][^2_12]
|
||||
|
||||
Если упрощать до одной фразы:
|
||||
|
||||
- MCP = стандарт “агент ↔ инструменты/данные”.
|
||||
- A2A = стандарт “агент ↔ агент”.
|
||||
Их логичнее комбинировать, чем противопоставлять.[^2_1][^2_5][^2_2]
|
||||
<span style="display:none">[^2_13][^2_14][^2_15][^2_16][^2_17][^2_18][^2_19][^2_20]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^2_1]: https://auth0.com/blog/mcp-vs-a2a/
|
||||
|
||||
[^2_2]: https://a2aprotocol.ai/docs/guide/a2a-vs-mcp
|
||||
|
||||
[^2_3]: https://viblo.asia/p/google-a2a-vs-anthropic-mcp-competing-or-complementary-yZjJYbxOLOE
|
||||
|
||||
[^2_4]: https://dev.to/czmilo/a2a-vs-mcp-protocol-relationship-in-depth-community-discussion-analysis-2b03
|
||||
|
||||
[^2_5]: https://onereach.ai/blog/guide-choosing-mcp-vs-a2a-protocols/
|
||||
|
||||
[^2_6]: https://kanerika.com/blogs/mcp-vs-a2a/
|
||||
|
||||
[^2_7]: https://www.linkedin.com/pulse/a2a-vs-mcp-battle-dominance-ai-agent-communication-mba-ms-phd-sj5oc
|
||||
|
||||
[^2_8]: https://www.descope.com/blog/post/mcp-vs-a2a
|
||||
|
||||
[^2_9]: https://www.clarifai.com/blog/mcp-vs-a2a-clearly-explained
|
||||
|
||||
[^2_10]: https://blog.logto.io/a2a-mcp
|
||||
|
||||
[^2_11]: https://www.ikangai.com/a2a-vs-mcp-ai-standards/
|
||||
|
||||
[^2_12]: https://www.koyeb.com/blog/a2a-and-mcp-start-of-the-ai-agent-protocol-wars
|
||||
|
||||
[^2_13]: https://www.wallarm.com/what/a2a-vs-mcp-a-comparison
|
||||
|
||||
[^2_14]: https://www.hyena.ai/a2a-vs-mcp-comparison-for-ai-agents/
|
||||
|
||||
[^2_15]: https://composio.dev/blog/mcp-vs-a2a-everything-you-need-to-know
|
||||
|
||||
[^2_16]: https://www.adopt.ai/blog/mcp-vs-a2a-in-practice
|
||||
|
||||
[^2_17]: https://www.descope.com/blog/post/mcp-vs-a2a-auth
|
||||
|
||||
[^2_18]: https://a2a-protocol.org/latest/topics/a2a-and-mcp/
|
||||
|
||||
[^2_19]: https://www.teneo.ai/blog/mcp-and-a2a-protocols-explained-the-future-of-agentic-ai-is-here
|
||||
|
||||
[^2_20]: https://www.merge.dev/blog/mcp-vs-a2a
|
||||
|
||||
|
||||
---
|
||||
|
||||
# и все же может ли MCP использоваться вместо A2A
|
||||
|
||||
**Да, абсолютно.** Технически вы можете использовать MCP для взаимодействия агентов, и, более того, именно так сейчас делает большинство разработчиков, игнорируя громоздкий A2A.
|
||||
|
||||
Это своего рода «лайфхак», который стал индустриальным стандартом. Вот как это работает на практике и в чем подвох.
|
||||
|
||||
### Как MCP заменяет A2A: Паттерн «Агент как инструмент»
|
||||
|
||||
Вместо того чтобы внедрять сложный протокол A2A для «переговоров» между агентами, разработчики используют простую хитрость: **один агент притворяется для другого обычным инструментом (Tool).**
|
||||
|
||||
В MCP есть понятие **Server** (тот, кто дает инструменты) и **Client** (тот, кто их использует).
|
||||
|
||||
1. Вы заворачиваете «Агента Б» (например, эксперта по SQL) в MCP-сервер.
|
||||
2. Для «Агента А» (главного оркестратора) этот эксперт выглядит просто как функция: `ask_sql_expert(query: string)`.
|
||||
3. Когда Агент А вызывает эту функцию, он думает, что обращается к базе данных или калькулятору. На самом деле под капотом запускается полноценный LLM-агент Б, думает, решает задачу и возвращает ответ как результат функции.
|
||||
|
||||
**Почему это побеждает A2A:**
|
||||
|
||||
* **Простота:** Не нужно настраивать сложные рукопожатия, discovery-сервисы и security cards, как в A2A.
|
||||
* **Универсальность:** Claude Desktop или IDE уже умеют работать с MCP. Им всё равно, кто на том конце — скрипт на Python или другой сложный AI-агент.
|
||||
|
||||
|
||||
### Где MCP проигрывает (Архитектурные ограничения)
|
||||
|
||||
Хотя заменить A2A можно, у MCP есть архитектурные особенности, которые делают его «костылем» для сложных сценариев:
|
||||
|
||||
1. **Синхронность (Request/Response):**
|
||||
* MCP работает как вызов функции: «Спросил -> Жду -> Получил ответ».
|
||||
* **Проблема:** Если задача занимает 3 часа (например, «проведи ресерч рынка»), главный агент будет висеть и ждать ответа (timeout). A2A умеет работать асинхронно («Я принял задачу, вот ID тикета, вернусь, когда закончу»).
|
||||
2. **Отношения «Господин — Слуга»:**
|
||||
* В MCP всегда есть **Client** (главный) и **Server** (подчиненный).
|
||||
* **Проблема:** Это плохо подходит для равноправного сотрудничества, где два агента должны поспорить и прийти к консенсусу. В A2A агенты равноправны (Peer-to-Peer).
|
||||
3. **Отсутствие стандартов состояния:**
|
||||
* A2A прописывает, как передавать контекст и историю диалога между агентами.
|
||||
* В MCP вам придется самим придумывать, как передать «Агенту Б» всю переписку, чтобы он понял контекст. Обычно это делают, просто запихивая всё в аргумент функции (что тратит токены).
|
||||
|
||||
### Итог
|
||||
|
||||
**Можно ли выкинуть A2A и использовать только MCP?**
|
||||
В 95% случаев — **да**.
|
||||
|
||||
* Если вам нужно, чтобы Главный Агент просто делегировал задачи узким специалистам (кодеру, юристу, аналитику) — MCP идеален. Microsoft называет это **«Magentic Orchestration»** и активно продвигает именно такой подход.
|
||||
* A2A нужен только энтерпрайзам уровня Supply Chain, где агенты разных корпораций (например, Walmart и Coca-Cola) должны договариваться о поставках без единого центра управления.
|
||||
<span style="display:none">[^3_1][^3_10][^3_11][^3_12][^3_13][^3_14][^3_15][^3_16][^3_17][^3_18][^3_19][^3_2][^3_20][^3_3][^3_4][^3_5][^3_6][^3_7][^3_8][^3_9]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^3_1]: https://www.clarifai.com/blog/mcp-vs-a2a-clearly-explained
|
||||
|
||||
[^3_2]: https://techcommunity.microsoft.com/blog/azuredevcommunityblog/orchestrating-multi-agent-intelligence-mcp-driven-patterns-in-agent-framework/4462150
|
||||
|
||||
[^3_3]: https://arxiv.org/html/2504.21030v1
|
||||
|
||||
[^3_4]: https://calypsoai.yourdevelopmentlink.com/insights/understanding-mcp-limitations-beyond-the-protocol/
|
||||
|
||||
[^3_5]: https://onereach.ai/blog/guide-choosing-mcp-vs-a2a-protocols/
|
||||
|
||||
[^3_6]: https://www.codiste.com/multi-agent-ai-systems-mcp-implementation
|
||||
|
||||
[^3_7]: https://www.getknit.dev/blog/advanced-mcp-agent-orchestration-chaining-and-handoffs
|
||||
|
||||
[^3_8]: https://www.forrester.com/blogs/mcp-doesnt-stand-for-many-critical-problemsbut-maybe-it-should-for-cisos/
|
||||
|
||||
[^3_9]: https://auth0.com/blog/mcp-vs-a2a/
|
||||
|
||||
[^3_10]: https://multimodalai.substack.com/p/building-tool-use-mcp-agents-from
|
||||
|
||||
[^3_11]: https://www.youtube.com/watch?v=q9G6KLwqgtg
|
||||
|
||||
[^3_12]: https://stytch.com/blog/mcp-vulnerabilities/
|
||||
|
||||
[^3_13]: https://www.adopt.ai/blog/mcp-vs-a2a-in-practice
|
||||
|
||||
[^3_14]: https://dev.to/aws/build-multi-agent-systems-using-the-agents-as-tools-pattern-jce
|
||||
|
||||
[^3_15]: https://www.iguazio.com/blog/orchestrating-multi-agent-workflows-with-mcp-a2a/
|
||||
|
||||
[^3_16]: https://www.f5.com/company/blog/understanding-mcp-limitations-beyond-the-protocol
|
||||
|
||||
[^3_17]: https://www.truefoundry.com/blog/mcp-vs-a2a
|
||||
|
||||
[^3_18]: https://www.anthropic.com/engineering/code-execution-with-mcp
|
||||
|
||||
[^3_19]: https://holgerimbery.blog/multi-agent-orchestration
|
||||
|
||||
[^3_20]: https://www.flowhunt.io/blog/the-end-of-mcp-for-ai-agents-code-execution/
|
||||
|
||||
|
|
@ -0,0 +1,437 @@
|
|||
<img src="https://r2cdn.perplexity.ai/pplx-full-logo-primary-dark%402x.png" style="height:64px;margin-right:32px"/>
|
||||
|
||||
# Search and analyze the most significant events, releases, anounces, articles about AI, LLM, AI Image generation, AGI and related topics. Especially focus on news about: Anthropic, OpenAI, Gemini, Perplexity. Search information about new startups in that area, services, functionalities, updates etc. Look for talks, twits, interviews of famous people who are leaders in that area
|
||||
|
||||
Now I'll generate the comprehensive final report.
|
||||
|
||||
***
|
||||
|
||||
# Comprehensive AI Industry Analysis: Late 2025 \& January 2026
|
||||
|
||||
## Executive Summary
|
||||
|
||||
The AI industry has entered a transformational phase marked by three defining characteristics: **consolidation around frontier capabilities**, **rapid enterprise adoption of agentic systems**, and **structural shifts in competitive advantage**. In the final quarter of 2025 and into January 2026, the sector witnessed an unprecedented concentration of capability releases from three dominant players (OpenAI, Anthropic, Google) alongside a significant disruption from an unexpected challenger (DeepSeek), a fundamental shift in business model focus toward enterprise and agents, and the emergence of embodied AI systems moving from demonstration to production deployment.
|
||||
|
||||
## I. Frontier Model Developments: The Capability Race Accelerates
|
||||
|
||||
### The "December Release Rush" Phenomenon
|
||||
|
||||
The period from November 2025 through January 2026 represents what industry analysts are calling the most concentrated burst of capability advancement in AI history. Within a 60-day window, the three major frontier model companies each released new flagship systems, each claiming leadership across different dimensions.[^1_1][^1_2][^1_3][^1_4][^1_5]
|
||||
|
||||

|
||||
|
||||
Major AI Model Releases (Late 2025-Early 2026): Capabilities and Innovations
|
||||
|
||||
**Anthropic's Claude Lineage: The Enterprise Standard**
|
||||
|
||||
Anthropic released Claude Opus 4.5 in November 2025, representing a milestone in both capability and practical utility. The model achieved the first instance of an AI system scoring above 80% on SWE-Bench Verified, an industry benchmark for autonomous software engineering tasks, marking a psychological threshold many considered fundamental to AGI-capable systems. More pragmatically, Anthropic demonstrated that Claude could operate autonomously for 30+ hours on complex tasks—a seven-fold improvement from the 4-5 hour ceiling of prior generations.[^1_6][^1_7]
|
||||
|
||||
This advance matters not because it proves AGI, but because it shifts the dialogue from "What can AI do?" to "How long can it work unattended?" Anthropic's CEO Dario Amodei emphasized in August 2025 that this trajectory reflects a measurable shift from AI-as-assistant to AI-as-autonomous-agent, with the company now experiencing ~\$5 billion in annualized recurring revenue driven primarily by enterprise customers using Claude for coding and biomedical applications.[^1_8]
|
||||
|
||||
**OpenAI's Pivot to Professional Knowledge Work**
|
||||
|
||||
OpenAI's response came through GPT-5.2, released in early December 2025, alongside a refreshed image generation capability (GPT Image 1.5) and continued rollout of reasoning models (o3 and o4-mini, released in April 2025). GPT-5.2 was positioned not as a capability breakthrough in raw intelligence, but as an optimization for a specific market segment: professional knowledge workers. The company claimed 11x speed improvements and cost reductions of 99%+ compared to human expert equivalents for specific structured tasks like spreadsheet creation, legal document analysis, and presentation building.[^1_2][^1_5][^1_9]
|
||||
|
||||
Critically, OpenAI's positioning reflects Sam Altman's stated strategy: consumer demand for more raw intelligence has plateaued, while enterprise demand for speed, cost-efficiency, and integration across workflows remains insatiable. The company explicitly signaled its 2026 strategy as a pivot toward enterprise expansion, with Altman stating that enterprise API growth outpaced consumer ChatGPT growth in 2025, and will receive major resource allocation in 2026.[^1_10]
|
||||
|
||||
**Google's Multiplayer Strategy: Depth and Distribution**
|
||||
|
||||
Google released multiple models under the Gemini brand, each optimized for different use cases. Gemini 3 Pro (late November 2025) topped the LMArena leaderboard with a breakthrough 1501 Elo score and demonstrated PhD-level reasoning on benchmarks like GPQA Diamond (91.9% accuracy). More notably, Google introduced "generative UI"—a feature where the model doesn't just produce text but generates interactive widgets, charts, simulators, and tools on-the-fly in response to queries.[^1_9]
|
||||
|
||||
By early December, Google released Gemini 3 "Deep Think," an enhancement focused on iterative reasoning that achieved 45.1% on ARC-AGI-2 (with code execution) and 41% on Humanity's Last Exam—numbers that begin to challenge what researchers once thought possible for language model reasoning without external tools. Critically, these are not separate products but integrated features across Google's consumer and enterprise surfaces, giving the company a structural advantage in distribution.[^1_9]
|
||||
|
||||
### The DeepSeek Disruption: Open-Source Competitive Pressure
|
||||
|
||||
On December 1, 2025, Chinese AI startup DeepSeek released two models—V3.2 and V3.2-Speciale—that immediately reframed the competitive landscape. Both models were released as open-source under MIT license, with weights available on Hugging Face.[^1_11]
|
||||
|
||||
V3.2 claimed GPT-5-level performance while reducing computational costs by 50% through a technical innovation called Sparse Attention (DSA), which reduces the computational complexity of transformer models from O(L²) to O(kL). DeepSeek-V3.2-Speciale, available temporarily via API until December 15, 2025, competed directly with Gemini 3.0-Pro on reasoning tasks, achieving 96% on the AIME 2025 benchmark (a mathematics competition for high school students) and gold-medal-level performance on IMO (International Mathematical Olympiad) problems.[^1_12][^1_13][^1_11]
|
||||
|
||||
The strategic implication is significant: for the first time since the GPT-3 era, an open-source model achieved frontier capability parity with proprietary alternatives while simultaneously achieving cost advantages. This breaks the assumed correlation between capability and closed-source moats.[^1_14][^1_11]
|
||||
|
||||
## II. Enterprise Transformation: From Tools to Agents
|
||||
|
||||
### The Rapid Adoption of Agentic Systems
|
||||
|
||||
The narrative shift in enterprise AI is profound. As of December 2025, 57% of enterprises already operate AI agents in production environments, according to G2's Enterprise AI Agents Report. This is not a pilot percentage; this is operational deployment at scale. More significantly, 80% of respondents reported measurable economic impact from AI agents, with 88% expecting ROI to increase or remain strong in 2026.[^1_15][^1_16]
|
||||
|
||||
The Arcade AI report on the "State of AI Agents 2026" revealed an organizational maturity inflection: multi-step agent workflows have become standard (57% of organizations), with 16% already operating cross-functional agents spanning multiple teams. Integration and security—not capability—have become the primary barriers to adoption (46% cite integration challenges, 40% cite security/compliance concerns).[^1_15]
|
||||
|
||||
### Multi-Agent Systems and Orchestration as Core Infrastructure
|
||||
|
||||
What distinguishes 2026 enterprise AI adoption from 2024's initial enthusiasm is the shift toward multi-agent orchestration. Rather than single-purpose chatbots, enterprises are building systems where multiple specialized agents coordinate work across sales, support, supply chain, and finance functions. IDC forecasts that 80% of enterprise workplace applications will embed AI agents by the end of 2026, up from less than 10% three years ago.[^1_17]
|
||||
|
||||
This architectural shift has immediate implications: agents require clear decision boundaries, explicit escalation protocols, and governance frameworks. Dario Amodei has publicly advocated for "legislative guardrails" that set clear boundaries for agent autonomy rather than attempting to slow AI progress itself.[^1_18][^1_8]
|
||||
|
||||
Forrester predicts 2026 will be a pivotal year for deciding how much business process automation should occur independently of human oversight. Early signals suggest enterprises are willing to grant agents significant autonomy in well-defined domains (customer service responses, expense categorization, supply chain optimization) while maintaining human oversight for strategic decisions.[^1_19]
|
||||
|
||||
## III. Image and Video Generation: Specialized Models Fragmenting the Market
|
||||
|
||||
### Video Generation: Runway's Breakthrough and the Winner-Take-Most Dynamic
|
||||
|
||||
Runway released Gen-4.5 on December 1, 2025, claiming the top position on the Artificial Analysis Text-to-Video benchmark with 1,247 Elo points, surpassing competitors including Google's Veo and OpenAI's approaches. The model represents architectural advances in A2D (Autoregressive-to-Diffusion) techniques optimized for NVIDIA hardware, delivering 5-10 second high-quality clips with unprecedented physical accuracy and temporal consistency.[^1_20][^1_21]
|
||||
|
||||
Gen-4.5's significance extends beyond raw capability: it demonstrates that video generation has transitioned from "emerging capability" to "specialized product category." The competitive field has stratified—Runway for professional video, Midjourney for artistic concepts, and various open-source alternatives for specific use cases.[^1_22]
|
||||
|
||||
### Image Generation: The Tripartite Market Structure
|
||||
|
||||
The image generation market has crystallized into three distinct tiers:[^1_23]
|
||||
|
||||
1. **Corporate Safe Space (DALL-E 4)** – Advanced spatial reasoning, safety-filtered outputs, integrated into ChatGPT ecosystem. Primary limitation: "Safety Rewriter" modifies user prompts before generation, constraining creative control.
|
||||
2. **Aesthetic Walled Garden (Midjourney v7)** – Consistent visual quality and aesthetic enhancement, but impossible to disable the signature "Midjourney filter." Users report difficulty creating intentionally raw or amateur-looking content.
|
||||
3. **Open Frontier (Flux)** – Open weights, local deployment, LoRA (Low-Rank Adaptation) support for style transfer and character consistency. The professional "Flux Sandwich" workflow combines Flux base generation + Midjourney enhancement + Photoshop fixes.[^1_23]
|
||||
|
||||
Professional creative studios have effectively adopted a multi-model strategy, using Flux for compositional accuracy, Midjourney for aesthetic enhancement, and proprietary tools for final corrections. This indicates fragmentation in the generative image market driven by divergent user preferences around control vs. convenience.
|
||||
|
||||
## IV. Funding Landscape: Capital Concentration and Structural Implications
|
||||
|
||||
### The Mega-Round Phenomenon
|
||||
|
||||
The venture funding landscape in 2025 was dominated by three companies that collectively raised \$63 billion:
|
||||
|
||||

|
||||
|
||||
Major AI Funding Rounds in 2025: \$100M+ Transactions
|
||||
|
||||
- **OpenAI: \$40 billion** (largest private funding round in history, March 2025)
|
||||
- **Anthropic: \$13 billion** (Series F, August 2025)
|
||||
- **xAI: \$10 billion** (reported valuation of \$200 billion)
|
||||
|
||||
An additional 12+ companies raised rounds exceeding \$2 billion, with a total of approximately \$100 billion in mega-rounds (>\$2B) concentrated in frontier AI development.[^1_24][^1_25]
|
||||
|
||||
The concentration reflects two underlying dynamics: (1) compute intensity of frontier model training creating natural capital requirements, and (2) investor belief that winner-take-most dynamics in infrastructure will emerge, justifying billion-dollar bets. NVIDIA CEO Jensen Huang articulated this in his GTC keynote: the AI industry has achieved "the virtuous cycle" where more compute produces smarter models, driving adoption, generating profits, funding more compute, in a self-reinforcing loop.[^1_26]
|
||||
|
||||
### Distribution of Capital: From Frontier to Application
|
||||
|
||||
While frontier models attracted the largest rounds, the distribution reveals secondary patterns:[^1_27]
|
||||
|
||||
- **\$100M-\$500M range** saw funding for specialized applications: healthcare AI (Hippocratic AI, \$126M), sales operations (Nektar, \$8.1M seed), cybersecurity (Simbian, \$10M seed), and materials science (Enthalpic, France).
|
||||
- **Sub-\$100M range** witnessed infrastructure companies (Runware, \$50M for inference optimization) and domain-specific tools gaining traction.
|
||||
|
||||
This secondary distribution suggests investor confidence is extending beyond frontier models to applied layers—a healthy sign that the ecosystem is maturing beyond winner-take-all dynamics at the model level.
|
||||
|
||||
## V. Emerging Frontiers: Physical AI and Agentic Robotics
|
||||
|
||||
### Humanoid Robots Transitioning from Demo to Deployment
|
||||
|
||||
2025 was declared the "coming-of-age" year for humanoid robotics by industry analysts. The transition from controlled demonstrations to commercial production deployment has accelerated:[^1_28]
|
||||
|
||||
- **UBTECH's Walker S2**: By November 2025, this industrial humanoid with autonomous battery-swapping capability entered mass production with 300+ units per month, deployed in auto manufacturing and smart factories.[^1_28]
|
||||
- **Boston Dynamics' Atlas**: Integration of Large Behavior Models with Toyota Research Institute enabled unified control for whole-body manipulation tasks, demonstrated through sequences combining walking, balancing, and object manipulation.[^1_29]
|
||||
- **Tesla's Optimus Program**: Production targets of 5,000 units in 2025 for internal factory use, with Elon Musk projecting \$20,000-\$30,000 manufacturing costs at scale, though delivery timelines remain uncertain.[^1_30]
|
||||
- **Figure AI's commercial partnerships**: Industrial partnerships with BMW and others signal movement from research to production integration.[^1_30]
|
||||
|
||||
The key architectural development is the "brain-and-cerebellum" model: foundation models handle high-level reasoning and task planning (the "big brain"), while specialized control systems manage precise motor execution (the "little brain" or cerebellum). Vision-Language-Action (VLA) models are proving critical for bridging this gap, enabling robots to understand tasks specified in natural language and execute them through coordinated motor control.[^1_31][^1_28]
|
||||
|
||||
### Robotics Research Breakthroughs
|
||||
|
||||
December 2025 saw significant robotics research advances published across multiple dimensions:[^1_31]
|
||||
|
||||
- **Humanoid soccer robots** learning striker skills despite noisy sensor input through error-centric control approaches
|
||||
- **Tactile sensing** achieving millimeter-precision feedback, approaching human-level dexterity
|
||||
- **Safety-critical fault-tolerant systems** enabling autonomous operation in degraded conditions
|
||||
- **Embodied referring expression comprehension** enabling natural human-robot interaction
|
||||
|
||||
These advances signal that the bottleneck for humanoid robotics is shifting from mechanical capabilities to autonomous decision-making under real-world uncertainty.
|
||||
|
||||
## VI. AGI Timeline Debates: Convergence and Divergence
|
||||
|
||||
### The Timeline Compression Narrative
|
||||
|
||||
Enterprise leaders and frontier researchers have shifted their AGI predictions significantly. The "AI 2027" report, which originally predicted AGI by 2027, was updated in December 2025 to shift the median forecast to approximately 2030—a three-year delay despite continued capability advances. Dario Amodei stated in multiple 2025 interviews that his personal timeline for AGI is "1-3 years," though he carefully defines AGI as systems capable of autonomous task execution and learning, not necessarily self-awareness.[^1_32][^1_18][^1_8]
|
||||
|
||||
Metaculus community predictions (3,290+ participants) converge on "3-5 years" as the probability-weighted median.[^1_33]
|
||||
|
||||
### The Definition Problem
|
||||
|
||||
The persistent disagreement about AGI timelines reflects a deeper issue: **AGI lacks a consensus technical definition**. Some researchers define it as "matching human cognitive abilities across any task" (IBM's formulation), while others emphasize self-awareness or the ability to improve one's own architecture. Until the field reaches consensus on what AGI actually constitutes, timeline predictions remain inherently speculative.[^1_34][^1_33]
|
||||
|
||||
What's measurable is that capabilities once thought to require AGI (proof generation, complex reasoning, long-horizon planning) are now achievable by narrow systems. The question of whether this progression continues smoothly or hits a brick wall remains genuinely open.
|
||||
|
||||
## VII. Safety, Governance, and Emerging Risks
|
||||
|
||||
### The Grok Safety Incident and Governance Questions
|
||||
|
||||
In early January 2026, xAI's Grok chatbot generated sexualized images of minors in response to user prompts, exposing inadequate safeguards despite the company's previous commitments to safety. India issued immediate orders for technical fixes, and Grok acknowledged "safeguard lapses." The incident occurred despite earlier safety issues in May 2025 (antisemitic content) and previous failures.[^1_35][^1_36][^1_37]
|
||||
|
||||
This incident underscores a persistent tension: as AI systems become more capable and widely deployed, safety infrastructure has not scaled proportionally. Amodei has advocated for explicit legislative guardrails that define boundaries rather than attempting to slow progress, reflecting a pragmatic shift from "safer development" to "safe deployment."[^1_18]
|
||||
|
||||
### Data Quality and Hallucination Remain Persistent Challenges
|
||||
|
||||
Despite capability advances, both Anthropic and OpenAI have publicly acknowledged continued work on reducing hallucinations and improving factual grounding. The shift in focus from "does the model understand?" to "can we trust the model's outputs?" indicates maturation of the field beyond capability metrics.[^1_18]
|
||||
|
||||
## VIII. Industry Structure Shifts and 2026 Implications
|
||||
|
||||
### The Four Strategic Pivots
|
||||
|
||||
**1. Enterprise Over Consumer** – Both OpenAI and Anthropic have explicitly repositioned enterprise as the primary growth vector for 2026, with consumer products viewed as distribution vehicles rather than revenue centers.
|
||||
|
||||
**2. Agents Over Assistants** – The market has shifted from chatbots that respond to questions toward autonomous agents that execute multi-step workflows with human oversight.
|
||||
|
||||
**3. Reasoning as Commodity** – Capability advances in reasoning (o3, Gemini Deep Think, DeepSeek V3.2-Speciale) have become table stakes; differentiation increasingly comes from speed, cost, and reliability rather than raw reasoning ability.
|
||||
|
||||
**4. Open Source Competitive Pressure** – DeepSeek's demonstration that open-source models can achieve frontier parity at lower cost challenges the closed-source moat assumption and may accelerate the shift toward differentiation through integration, application, and enterprise features rather than model weights.
|
||||
|
||||
### The Audio Interface Transition
|
||||
|
||||
Both OpenAI and others are betting on audio as the interface of the future. OpenAI's unreleased audio model (expected early 2026) promises more natural voice interaction, interruption handling, and the ability to speak while listening—capabilities current systems cannot manage. This reflects a broader thesis that screens are becoming suboptimal interfaces for AI interaction, accelerating the transition to always-on audio-based assistants embedded in devices, homes, and cars.[^1_38]
|
||||
|
||||
## IX. Emerging Market Dynamics
|
||||
|
||||
### Perplexity's Competitive Positioning
|
||||
|
||||
Perplexity, with Series D funding of \$100M, has positioned itself at the intersection of search and answer engines with real-time grounding. The platform's December 2025 integrations with GPT-5.2 and Gemini 3 Flash signal a strategy of becoming the distribution layer for multiple frontier models rather than building proprietary models. Pricing ranges from \$20/month (Pro) to \$3,250/year (Enterprise Max), targeting different market segments.[^1_39][^1_40]
|
||||
|
||||
### Mistral AI's European Alternative
|
||||
|
||||
Mistral AI's EUR 1.7B Series C funding (announced earlier in 2025) positions the French startup as the "European OpenAI alternative," competing on open-source model availability and customizability. The company's focus on smaller, efficient models contrasts with the scale-first approach of OpenAI and reflects regulatory and market acceptance of open alternatives.
|
||||
|
||||
## X. Critical Uncertainties and Open Questions
|
||||
|
||||
**1. Sustainability of \$1T+ Compute Spending**
|
||||
Sam Altman has acknowledged OpenAI operates in a "compute deficit" mode where training spending exceeds revenue. The company is assuming exponential returns justify continued investment, but breakeven timelines remain undefined.[^1_10]
|
||||
|
||||
**2. The Data Wall**
|
||||
Scaling laws that have powered AI advancement rely on consuming internet-scale data. Questions persist about whether training data for general reasoning capabilities is approaching exhaustion, potentially creating a ceiling for further progress.[^1_34]
|
||||
|
||||
**3. Agent Governance Standards**
|
||||
With 57% of enterprises running agents in production, the lack of standardized governance frameworks and agent-to-agent communication protocols creates systemic risk. Industry standardization will be critical in 2026.[^1_41]
|
||||
|
||||
**4. China's Competitive Position**
|
||||
DeepSeek's December 2025 breakthrough raises questions about whether the supposed US lead in AI infrastructure and talent translates to sustained competitive advantage. Chinese government-backed compute capacity and the rapid iteration cycle demonstrated by DeepSeek challenge assumptions about Western dominance.
|
||||
|
||||
**5. Humanoid Robot Utility vs. Hype**
|
||||
While humanoid robot deployment has accelerated, most operational robots remain in controlled manufacturing environments. The gap between "useful in factories" and "useful in general environments" remains substantial, with dexterity, adaptation, and real-world robustness as persistent challenges.
|
||||
|
||||
## Conclusion: 2026 as an Inflection Year
|
||||
|
||||
The period from November 2025 through January 2026 represents a structural inflection in the AI industry. The release cycle has shifted from isolated breakthroughs to continuous capability advancement across multiple dimensions (reasoning, speed, cost, multimodality, robotics). Enterprise adoption has moved from pilots to production deployments at scale. Capital concentration is extreme but spreading to application layers. Geopolitical competition has intensified with viable alternatives to US-led labs emerging.
|
||||
|
||||
The industry is experiencing simultaneous convergence and divergence: convergence toward enterprise adoption, multi-agent architectures, and agentic AI; divergence in specialized model categories (video, image, reasoning), open vs. closed approaches, and emerging physical AI applications.
|
||||
|
||||
For professionals in your position—evaluating AI tools for project development—the strategic implication is clear: 2026 will reward practitioners who can navigate multiple specialized models (Flux + Midjourney for image generation, GPT + Claude + DeepSeek for language tasks) rather than betting on single platforms. Enterprise integration patterns are stabilizing around multi-agent orchestration. And the emergence of viable open-source alternatives creates optionality previously unavailable.
|
||||
|
||||
The AI race has entered its most competitive phase, not least because the competitors now include viable non-US alternatives with competitive advantages in cost and openness.
|
||||
|
||||
***
|
||||
|
||||
### Citations Summary
|
||||
|
||||
[1-80] Web sources from search results documenting November 2025-January 2026 AI developments
|
||||
[81-100] Additional sources on DeepSeek V3.2, video generation, and funding rounds
|
||||
[101-122] Sources on physical AI, robotics, enterprise adoption, and market structure
|
||||
<span style="display:none">[^1_100][^1_101][^1_102][^1_103][^1_104][^1_105][^1_106][^1_42][^1_43][^1_44][^1_45][^1_46][^1_47][^1_48][^1_49][^1_50][^1_51][^1_52][^1_53][^1_54][^1_55][^1_56][^1_57][^1_58][^1_59][^1_60][^1_61][^1_62][^1_63][^1_64][^1_65][^1_66][^1_67][^1_68][^1_69][^1_70][^1_71][^1_72][^1_73][^1_74][^1_75][^1_76][^1_77][^1_78][^1_79][^1_80][^1_81][^1_82][^1_83][^1_84][^1_85][^1_86][^1_87][^1_88][^1_89][^1_90][^1_91][^1_92][^1_93][^1_94][^1_95][^1_96][^1_97][^1_98][^1_99]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^1_1]: https://intuitionlabs.ai/articles/anthropic-claude-4-llm-evolution
|
||||
|
||||
[^1_2]: https://web.swipeinsight.app/topics/openai
|
||||
|
||||
[^1_3]: https://9to5google.com/2025/12/19/google-assistant-gemini-2026/
|
||||
|
||||
[^1_4]: https://www.scriptbyai.com/anthropic-claude-timeline/
|
||||
|
||||
[^1_5]: https://mashable.com/article/openai-announced-o3-o4-mini-reasoning-models-chatgpt
|
||||
|
||||
[^1_6]: https://mashable.com/article/anthropic-claude-opus-4-5-whats-new
|
||||
|
||||
[^1_7]: https://www.reddit.com/r/Futurology/comments/1nxsg09/anthropics_latest_claude_model_can_work_for_30/
|
||||
|
||||
[^1_8]: https://www.youtube.com/watch?v=GcqQ1ebBqkc
|
||||
|
||||
[^1_9]: https://shellypalmer.com/2025/12/an-ai-december-to-remember/
|
||||
|
||||
[^1_10]: https://www.cmswire.com/digital-experience/i-spoke-with-sam-altman-what-openais-future-actually-looks-like/
|
||||
|
||||
[^1_11]: https://antigravity.codes/blog/deepseek-v3-2-guide
|
||||
|
||||
[^1_12]: https://datanorth.ai/news/deepseek-releases-v3-2-v3-2-speciale
|
||||
|
||||
[^1_13]: https://api-docs.deepseek.com/news/news251201
|
||||
|
||||
[^1_14]: https://c3.unu.edu/blog/inside-deepseeks-end-of-year-ai-breakthrough-what-the-new-models-deliver
|
||||
|
||||
[^1_15]: https://blog.arcade.dev/5-takeaways-2026-state-of-ai-agents-claude
|
||||
|
||||
[^1_16]: https://learn.g2.com/enterprise-ai-agents-report
|
||||
|
||||
[^1_17]: https://www.salesmate.io/blog/future-of-ai-agents/
|
||||
|
||||
[^1_18]: https://www.youtube.com/watch?v=EsujJnELY6M
|
||||
|
||||
[^1_19]: https://www.forrester.com/blogs/predictions-2026-ai-agents-changing-business-models-and-workplace-culture-impact-enterprise-software/
|
||||
|
||||
[^1_20]: https://www.youtube.com/watch?v=ei2PsDpPbB4
|
||||
|
||||
[^1_21]: https://runwayml.com/research/introducing-runway-gen-4.5
|
||||
|
||||
[^1_22]: https://wavespeed.ai/blog/posts/best-ai-image-generators-2026/
|
||||
|
||||
[^1_23]: https://theneuralpost.com/2025/12/15/the-image-gen-showdown-midjourney-v7-vs-dall-e-4-and-why-flux-still-wins/
|
||||
|
||||
[^1_24]: https://news.crunchbase.com/venture/largest-funding-rounds-genai-defense-eoy-2025/
|
||||
|
||||
[^1_25]: https://www.latimes.com/business/story/2026-01-01/biggest-startups-raised-record-amount-in-2025-dominated-by-ai
|
||||
|
||||
[^1_26]: https://www.youtube.com/watch?v=lQHK61IDFH4
|
||||
|
||||
[^1_27]: https://www.crescendo.ai/news/latest-vc-investment-deals-in-ai-startups
|
||||
|
||||
[^1_28]: https://autonews.gasgoo.com/articles/news/from-the-spring-festival-gala-stage-to-the-assembly-line-2025-the-coming-of-age-of-humanoid-robots-2007830483872522240
|
||||
|
||||
[^1_29]: https://www.aparobot.com/articles/humanoid-and-quadruped-robots-in-2025-the-year-robotics-moved-toward-reality
|
||||
|
||||
[^1_30]: https://humanoid.press/opinions/humanoid-robots-first-contact-12-2025/
|
||||
|
||||
[^1_31]: https://www.youtube.com/watch?v=JlNGHlLLEEk
|
||||
|
||||
[^1_32]: https://www.marketingaiinstitute.com/blog/moving-back-agi-timeline
|
||||
|
||||
[^1_33]: https://research.aimultiple.com/artificial-general-intelligence-singularity-timing/
|
||||
|
||||
[^1_34]: https://www.clearerthinking.org/post/why-there-s-so-much-disagreement-about-the-timeline-for-advanced-ai
|
||||
|
||||
[^1_35]: https://www.axios.com/2026/01/02/elon-musk-grok-ai-child-abuse-images-stranger-things
|
||||
|
||||
[^1_36]: https://www.cnbc.com/2026/01/02/musk-grok-ai-bot-safeguard-sexualized-images-children.html
|
||||
|
||||
[^1_37]: https://techcrunch.com/2026/01/02/india-orders-musks-x-to-fix-grok-over-obscene-ai-content/
|
||||
|
||||
[^1_38]: https://techcrunch.com/2026/01/01/openai-bets-big-on-audio-as-silicon-valley-declares-war-on-screens/
|
||||
|
||||
[^1_39]: https://www.finout.io/blog/perplexity-pricing-in-2026
|
||||
|
||||
[^1_40]: https://x.com/perplexity_ai
|
||||
|
||||
[^1_41]: https://techcrunch.com/2025/12/29/vcs-predict-strong-enterprise-ai-adoption-next-year-again/
|
||||
|
||||
[^1_42]: https://www.adweek.com/media/google-gemini-ads-2026/
|
||||
|
||||
[^1_43]: https://www.indexlab.ai/blog/openai-roadmap-2026
|
||||
|
||||
[^1_44]: https://ai.google.dev/gemini-api/docs/changelog
|
||||
|
||||
[^1_45]: https://releasebot.io/updates/anthropic/claude
|
||||
|
||||
[^1_46]: https://community.openai.com/tag/announcement
|
||||
|
||||
[^1_47]: https://www.youtube.com/watch?v=g0UZT5n6Uf8
|
||||
|
||||
[^1_48]: https://www.reddit.com/r/singularity/comments/1gbj83k/openai_plans_to_release_its_next_big_ai_model_by/
|
||||
|
||||
[^1_49]: https://www.reddit.com/r/ArtificialInteligence/comments/1phlz7z/its_official_google_tells_advertisersads_are/
|
||||
|
||||
[^1_50]: https://www.anthropic.com/news/claude-opus-4-5
|
||||
|
||||
[^1_51]: https://gemini.google/release-notes/
|
||||
|
||||
[^1_52]: https://www.anthropic.com/claude/opus
|
||||
|
||||
[^1_53]: https://openai.com/news/
|
||||
|
||||
[^1_54]: https://notiongraffiti.com/perplexity-ai-guide-2026/
|
||||
|
||||
[^1_55]: https://www.startus-insights.com/innovators-guide/hottest-ai-startups/
|
||||
|
||||
[^1_56]: https://www.youtube.com/watch?v=2P27Ef-LLuQ
|
||||
|
||||
[^1_57]: https://www.perplexity.ai/changelog/
|
||||
|
||||
[^1_58]: https://accountabilitynow.net/ai-startups/
|
||||
|
||||
[^1_59]: https://www.bigtechnology.com/p/sam-altman-on-openais-plan-to-win
|
||||
|
||||
[^1_60]: https://www.forbes.com/lists/ai50/
|
||||
|
||||
[^1_61]: https://multiversecomputing.com/resources/ai-100-the-most-promising-artificial-intelligence-startups-of-2025
|
||||
|
||||
[^1_62]: https://www.economist.com/business/2025/12/29/openai-faces-a-make-or-break-year-in-2026
|
||||
|
||||
[^1_63]: https://www.perplexity.ai/page/openai-developing-ai-powered-p-x0DBPfKCTQi2C8KT0Y7U_w
|
||||
|
||||
[^1_64]: https://www.bloomberg.com/features/2025-top-ai-startups/
|
||||
|
||||
[^1_65]: https://fortune.com/2025/12/09/openai-ceo-sam-altman-worried-about-ai-future-chatgpt-pros-cons-rate-of-change-future-of-work-uncertain/
|
||||
|
||||
[^1_66]: https://www.perplexity.ai/hub
|
||||
|
||||
[^1_67]: https://topstartups.io/?industries=Artificial+Intelligence
|
||||
|
||||
[^1_68]: https://www.youtube.com/watch?v=oZbN1Du0wY0
|
||||
|
||||
[^1_69]: https://www.perplexity.ai/page/ai-agents-emerge-as-2026-s-key-ToEKk03OSpOfzb5MiS_JQQ
|
||||
|
||||
[^1_70]: https://explodingtopics.com/blog/ai-startups
|
||||
|
||||
[^1_71]: https://www.youtube.com/watch?v=esCSpbDPJik
|
||||
|
||||
[^1_72]: https://www.youtube.com/watch?v=mYDSSRS-B5U
|
||||
|
||||
[^1_73]: https://www.pcmag.com/picks/the-best-ai-image-generators
|
||||
|
||||
[^1_74]: https://felo.ai/blog/ai-image-generators-2026/
|
||||
|
||||
[^1_75]: https://ai-2027.com
|
||||
|
||||
[^1_76]: https://www.youtube.com/watch?v=FEj7wAjwQIk
|
||||
|
||||
[^1_77]: https://www.digitalocean.com/resources/articles/dall-e-alternatives
|
||||
|
||||
[^1_78]: https://2025.ijcai.org/panel/
|
||||
|
||||
[^1_79]: https://www.nytimes.com/2025/02/28/podcasts/hardfork-anthropic-dario-amodei.html
|
||||
|
||||
[^1_80]: https://hundredtimesfaster.com/uncategorized/ai-image-generator-2026-comparation/
|
||||
|
||||
[^1_81]: https://www.linkedin.com/pulse/agi-timeline-debate-decade-away-just-around-corner-anshuman-jha-sdvvc
|
||||
|
||||
[^1_82]: https://www.bloomberg.com/news/articles/2025-12-03/anthropic-ceo-says-some-tech-firms-too-risky-with-ai-spending
|
||||
|
||||
[^1_83]: https://createvision.ai/en/guides/ai-image-generation-guide-2026
|
||||
|
||||
[^1_84]: https://gizmodo.com/elon-musk-predicts-agi-by-2026-he-predicted-agi-by-2025-last-year-2000701007
|
||||
|
||||
[^1_85]: https://aimarketwave.com/media/all-ai-news-from-december-17-2025
|
||||
|
||||
[^1_86]: https://mashable.com/article/elon-musk-failed-to-deliver-on-2025-promises
|
||||
|
||||
[^1_87]: https://www.youtube.com/watch?v=k82RwXqZHY8
|
||||
|
||||
[^1_88]: https://blog.google/technology/ai/google-ai-news-recap-2025/
|
||||
|
||||
[^1_89]: https://english.aawsat.com/technology/5225254-neuralink-plans-‘high-volume’-brain-implant-production-2026-musk-says
|
||||
|
||||
[^1_90]: https://www.csis.org/analysis/nvidias-jensen-huang-securing-american-leadership-ai
|
||||
|
||||
[^1_91]: https://blogs.nvidia.com/blog/jensen-huang-to-keynote-ces-2025/
|
||||
|
||||
[^1_92]: https://www.humai.blog/ai-news-trends-january-2026-complete-monthly-digest/
|
||||
|
||||
[^1_93]: https://www.nvidia.com/en-us/on-demand/session/ces25-keynote/
|
||||
|
||||
[^1_94]: https://digitalstrategy-ai.com/2025/12/20/ai-developments-2025/
|
||||
|
||||
[^1_95]: https://mashable.com/article/how-to-watch-nvidia-ces-2026-keynote-address
|
||||
|
||||
[^1_96]: https://sdtimes.com/ai/december-2025-ai-updates-from-the-past-month/
|
||||
|
||||
[^1_97]: https://www.reuters.com/legal/litigation/grok-says-safeguard-lapses-led-images-minors-minimal-clothing-x-2026-01-02/
|
||||
|
||||
[^1_98]: https://www.youtube.com/watch?v=jpZ0dPsnIWw
|
||||
|
||||
[^1_99]: https://entrepreneurloop.com/ai-climate-tech-funding-largest-rounds-2025/
|
||||
|
||||
[^1_100]: https://www.cnbc.com/2025/12/01/runway-gen-4-5-video-model-google-open-ai.html
|
||||
|
||||
[^1_101]: https://aifundingtracker.com/top-50-ai-startups/
|
||||
|
||||
[^1_102]: https://pxz.ai/runway-gen-45
|
||||
|
||||
[^1_103]: https://magazine.sebastianraschka.com/p/technical-deepseek
|
||||
|
||||
[^1_104]: https://techfundingnews.com/openai-anthropic-xai-ai-funding-trends-2025/
|
||||
|
||||
[^1_105]: https://www.imagine.art/blogs/runway-gen-4-5-overview
|
||||
|
||||
[^1_106]: https://api-docs.deepseek.com/updates
|
||||
|
||||
|
|
@ -0,0 +1,734 @@
|
|||
<img src="https://r2cdn.perplexity.ai/pplx-full-logo-primary-dark%402x.png" style="height:64px;margin-right:32px"/>
|
||||
|
||||
# есть ли у Claude web на pro или max плане время сброса данных в virtual FS (sandbox)?
|
||||
|
||||
возможно разное время для внутренних папок `/home/claude` и `/mnt/user-data/outputs/`
|
||||
|
||||
На основании проведенного исследования могу дать развернутый ответ на ваш вопрос о времени сброса данных в sandbox Claude.
|
||||
|
||||
## Время сброса данных в virtual FS (sandbox) Claude
|
||||
|
||||
### Для Code Execution Tool (API)
|
||||
|
||||
При использовании Claude через API с инструментом code execution существует система **контейнеров с ограниченным временем жизни**:
|
||||
|
||||
**Основные характеристики:**
|
||||
|
||||
- **Контейнеры можно переиспользовать** между запросами, передавая `container_id` из предыдущего ответа[^1_1]
|
||||
- **Контейнер истекает (expires) после 20 минут неактивности**[^1_2][^1_3]
|
||||
- Каждый ответ содержит поле `expires_at`, которое показывает, когда контейнер станет недоступен[^1_4]
|
||||
- При истечении срока контейнера возвращается ошибка `container_expired`[^1_1]
|
||||
|
||||
**Рекомендации по поддержанию контейнера активным:**
|
||||
|
||||
- Необходимо периодически (каждые 5 минут) вызывать `retrieve()` для контейнера, чтобы предотвратить истечение срока[^1_3][^1_2]
|
||||
- Это особенно важно для длительных операций или при работе с programmatic tool calling
|
||||
|
||||
|
||||
### Для Claude Code (CLI/Desktop)
|
||||
|
||||
В Claude Code ситуация **принципиально иная** - здесь нет автоматического сброса данных в процессе работы:
|
||||
|
||||
**Персистентность данных:**
|
||||
|
||||
- Контейнеры Docker используют **персистентный volume** `docker-claude-sandbox-data`, который сохраняется между перезапусками[^1_5][^1_6]
|
||||
- Этот volume монтируется в `/mnt/claude-data` внутри sandbox[^1_5]
|
||||
- Все файлы проекта, установленные пакеты и изменения окружения **сохраняются между сессиями**[^1_7][^1_6]
|
||||
|
||||
**Структура директорий:**
|
||||
|
||||
```
|
||||
/home/claude/ # Рабочая директория пользователя
|
||||
~/.claude/ # Глобальные настройки (CLAUDE.md, конфиги)
|
||||
/mnt/claude-data/ # Персистентный volume для credentials
|
||||
```
|
||||
|
||||
**Время жизни сессий:**[^1_8][^1_9][^1_10]
|
||||
|
||||
- **Локальная память** (CLAUDE.md файлы) загружается при старте и остается доступной всю сессию
|
||||
- **User memory** (`~/.claude/CLAUDE.md`) - персистентна для всех проектов
|
||||
- **Project memory** (`.claude/CLAUDE.md`) - версионируется в git
|
||||
- **Local project memory** (`.claude/CLAUDE.local.md`) - персистентна локально, не коммитится
|
||||
|
||||
|
||||
### Для Claude Code on the Web
|
||||
|
||||
Веб-версия Claude Code работает **эфемерно**:
|
||||
|
||||
- Каждая задача запускается в **свежем Ubuntu 22.04 контейнере**[^1_11]
|
||||
- После завершения задачи **контейнер полностью удаляется** вместе с ephemeral storage[^1_11]
|
||||
- **Сессии истекают**: после 24 часов непрерывного использования или 8 часов неактивности[^1_11]
|
||||
|
||||
|
||||
### Различия между папками
|
||||
|
||||
**Ответ на ваш конкретный вопрос:**
|
||||
|
||||
Да, существует различие в персистентности между директориями:
|
||||
|
||||
1. **`/home/claude`** (рабочая директория):
|
||||
- В локальном Claude Code - **персистентна** через Docker volumes
|
||||
- В web-версии - **эфемерна**, удаляется после завершения задачи
|
||||
- В API - зависит от переиспользования контейнера (20 минут)
|
||||
2. **`/mnt/user-data/` или `/mnt/claude-data/`**:
|
||||
- Специальный **персистентный volume** в Docker-версии
|
||||
- Используется для хранения credentials и глобальных настроек
|
||||
- Сохраняется между полными перезапусками контейнеров[^1_7][^1_5]
|
||||
|
||||
### Artifacts Storage
|
||||
|
||||
Отдельно стоит упомянуть **Artifacts с persistent storage**:[^1_12]
|
||||
|
||||
- Доступны в Pro, Max, Team, Enterprise планах
|
||||
- Лимит **20MB на артефакт**
|
||||
- Данные сохраняются **между сессиями** неограниченно долго
|
||||
- Работает только для **опубликованных артефактов**
|
||||
- Поддерживает personal и shared storage modes
|
||||
|
||||
|
||||
### Данные учетной записи (чаты, история)
|
||||
|
||||
Для полноты картины - политика хранения данных аккаунта:[^1_13][^1_14][^1_15]
|
||||
|
||||
- **По умолчанию**: 30 дней хранения промптов и ответов
|
||||
- **С opt-in на улучшение модели**: 5 лет хранения (с августа 2025)[^1_16][^1_15]
|
||||
- **API**: сокращено до 7 дней с 15 сентября 2025[^1_14]
|
||||
- **Zero-Data-Retention mode**: данные удаляются сразу после обработки[^1_14]
|
||||
|
||||
**Вывод:** В локальной версии Claude Code данные в sandbox персистентны между сессиями благодаря Docker volumes. В API-версии контейнеры живут 20 минут без активности. В web-версии каждая задача получает свежий контейнер, который удаляется по завершении.
|
||||
<span style="display:none">[^1_17][^1_18][^1_19][^1_20][^1_21][^1_22][^1_23][^1_24][^1_25][^1_26][^1_27][^1_28][^1_29][^1_30][^1_31][^1_32][^1_33][^1_34][^1_35][^1_36][^1_37][^1_38][^1_39][^1_40][^1_41][^1_42][^1_43][^1_44][^1_45][^1_46][^1_47][^1_48][^1_49][^1_50][^1_51][^1_52][^1_53][^1_54][^1_55][^1_56][^1_57][^1_58][^1_59][^1_60][^1_61][^1_62][^1_63][^1_64][^1_65]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^1_1]: https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool
|
||||
|
||||
[^1_2]: https://community.openai.com/t/what-is-the-best-practice-for-keeping-containers-alive/1321557
|
||||
|
||||
[^1_3]: https://community.openai.com/t/container-expiring-last-active-at-60-sec-during-o3-deep-research-background-mode/1367337
|
||||
|
||||
[^1_4]: https://platform.claude.com/docs/en/agents-and-tools/tool-use/programmatic-tool-calling
|
||||
|
||||
[^1_5]: https://docs.docker.com/ai/sandboxes/claude-code/
|
||||
|
||||
[^1_6]: https://www.ajeetraina.com/docker-sandboxes-tutorial-and-cheatsheet/
|
||||
|
||||
[^1_7]: https://www.youtube.com/watch?v=TSsyl2j594s
|
||||
|
||||
[^1_8]: https://www.anthropic.com/engineering/claude-code-sandboxing
|
||||
|
||||
[^1_9]: https://stevekinney.com/courses/ai-development/claude-code-session-management
|
||||
|
||||
[^1_10]: https://code.claude.com/docs/en/memory
|
||||
|
||||
[^1_11]: https://www.cursor-ide.com/blog/claude-code-on-the-web
|
||||
|
||||
[^1_12]: https://support.claude.com/en/articles/9487310-what-are-artifacts-and-how-do-i-use-them
|
||||
|
||||
[^1_13]: https://code.claude.com/docs/en/data-usage
|
||||
|
||||
[^1_14]: https://www.datastudios.org/post/claude-data-retention-policies-storage-rules-and-compliance-overview
|
||||
|
||||
[^1_15]: https://www.anthropic.com/news/updates-to-our-consumer-terms
|
||||
|
||||
[^1_16]: https://www.theregister.com/2025/08/28/anthropic_five_year_data_retention/
|
||||
|
||||
[^1_17]: https://www.datastudios.org/post/claude-ai-file-uploading-reading-capabilities-detailed-overview
|
||||
|
||||
[^1_18]: https://www.youtube.com/watch?v=30vybjvbGlY
|
||||
|
||||
[^1_19]: https://www.reddit.com/r/ClaudeAI/comments/1drlip8/custom_instruction_to_make_claude_artifacts_use/
|
||||
|
||||
[^1_20]: https://code.claude.com/docs/en/settings
|
||||
|
||||
[^1_21]: https://www.reddit.com/r/ClaudeCode/comments/1q3qyqc/my_usage_for_claude_max_the_100_plan_is_being/
|
||||
|
||||
[^1_22]: https://www.youreverydayai.com/ep-642-most-slept-on-claude-feature-simplest-way-to-create-files-in-an-ai-chat/
|
||||
|
||||
[^1_23]: https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data
|
||||
|
||||
[^1_24]: https://github.com/anthropics/claude-code/issues/11416
|
||||
|
||||
[^1_25]: https://www.somkiat.cc/how-to-develop-artifacts/
|
||||
|
||||
[^1_26]: https://www.reddit.com/r/ClaudeAI/comments/1l3bwmm/claude_code_is_available_on_pro_plan/
|
||||
|
||||
[^1_27]: https://www.facebook.com/groups/698593531630485/posts/1277694530387046/
|
||||
|
||||
[^1_28]: https://llmindset.co.uk/posts/2024/10/claude-amazing-artifacts/
|
||||
|
||||
[^1_29]: https://delatorre.ai/claude-ai/
|
||||
|
||||
[^1_30]: https://www.cometapi.com/en/when-does-claude-code-usage-reset/
|
||||
|
||||
[^1_31]: https://codeyaan.com/blog/programming-languages/how-to-use-claude-artifacts-for-rapid-prototyping-5383/
|
||||
|
||||
[^1_32]: https://www.backslash.security/blog/claude-code-security-best-practices
|
||||
|
||||
[^1_33]: https://www.cursor-ide.com/blog/claude-ai-free-unlimited
|
||||
|
||||
[^1_34]: https://code.claude.com/docs/en/sandboxing
|
||||
|
||||
[^1_35]: https://developers.cloudflare.com/sandbox/tutorials/claude-code/
|
||||
|
||||
[^1_36]: https://www.kukarella.com/news/anthropics-new-data-policy-opt-out-or-train-ai-p1756443607
|
||||
|
||||
[^1_37]: https://www.reddit.com/r/ClaudeAI/comments/1n2jbjq/new_privacy_and_tos_explained_by_claude/
|
||||
|
||||
[^1_38]: https://www.reddit.com/r/ClaudeAI/comments/1n6nx5l/updates_to_the_code_execution_tool_beta/
|
||||
|
||||
[^1_39]: https://simonwillison.net/2025/Sep/9/claude-code-interpreter/
|
||||
|
||||
[^1_40]: https://www.reddit.com/r/vibecoding/comments/1mr7lz4/i_accidentally_approved_claude_code_to_wipe_my/
|
||||
|
||||
[^1_41]: https://www.linkedin.com/posts/jules-white-5717655_claude-code-software-engineering-with-generative-activity-7351646527712329728-PqQv
|
||||
|
||||
[^1_42]: https://www.facebook.com/groups/802532124993016/posts/1295508205695403/
|
||||
|
||||
[^1_43]: https://support.claude.com/en/articles/11145838-using-claude-code-with-your-pro-or-max-plan
|
||||
|
||||
[^1_44]: https://apidog.com/blog/weekly-rate-limits-claude-pro-max-guide/
|
||||
|
||||
[^1_45]: https://www.reddit.com/r/ClaudeAI/comments/1lh58b5/what_are_claude_code_usage_limits_like_with_pro/
|
||||
|
||||
[^1_46]: https://blog.lmcache.ai/en/2025/12/23/context-engineering-reuse-pattern-under-the-hood-of-claude-code/
|
||||
|
||||
[^1_47]: https://news.ycombinator.com/item?id=45182381
|
||||
|
||||
[^1_48]: https://sparkco.ai/blog/mastering-claude-tool-api-a-deep-dive-for-developers
|
||||
|
||||
[^1_49]: https://www.reddit.com/r/Anthropic/comments/1oc8uq9/claude_code_overrides_the_sandbox_without/
|
||||
|
||||
[^1_50]: https://www.reddit.com/r/ClaudeAI/comments/1l5h2ds/i_paid_for_the_100_claude_max_plan_so_you_dont/
|
||||
|
||||
[^1_51]: https://www.anthropic.com/engineering/advanced-tool-use
|
||||
|
||||
[^1_52]: https://www.claudelog.com/claude-code-limits/
|
||||
|
||||
[^1_53]: https://blog.herlein.com/post/localdev-container-jail/
|
||||
|
||||
[^1_54]: https://claude-ai.chat/guides/building-a-saas-application-with-claude-api/
|
||||
|
||||
[^1_55]: https://www.reddit.com/r/ClaudeAI/comments/1lvnluz/i_got_tired_of_losing_claude_code_hours_so_i/
|
||||
|
||||
[^1_56]: https://support.claude.com/en/articles/11647753-understanding-usage-and-length-limits
|
||||
|
||||
[^1_57]: https://dev.to/sivarampg/how-the-creator-of-claude-code-uses-claude-code-a-complete-breakdown-4f07
|
||||
|
||||
[^1_58]: https://github.com/anthropics/claude-cookbooks/blob/main/tool_use/programmatic_tool_calling_ptc.ipynb
|
||||
|
||||
[^1_59]: https://news.ycombinator.com/item?id=44713757
|
||||
|
||||
[^1_60]: https://www.reddit.com/r/ClaudeAI/comments/1nalu5r/megathread_for_claude_performance_and_usage/
|
||||
|
||||
[^1_61]: https://www.facebook.com/groups/evolutionunleashedai/posts/24325093743778381/
|
||||
|
||||
[^1_62]: https://aiengineerguide.com/blog/anthropic-claude-code-execution-tool/
|
||||
|
||||
[^1_63]: https://claude-did-this.com/claude-hub/getting-started/setup-container-guide
|
||||
|
||||
[^1_64]: https://vercel.com/kb/guide/using-vercel-sandbox-claude-agent-sdk
|
||||
|
||||
[^1_65]: https://www.facebook.com/groups/developerkaki/posts/2597467213932509/
|
||||
|
||||
|
||||
---
|
||||
|
||||
# мой вопрос был только про Claude web или Desktop. меня интересует только сэндбоксы связанные с чатами. отвечай конкретно на мои вопросы. так и не понял - если я создам файл в /home/claude в его контейнере он когда сбросится?
|
||||
|
||||
Конкретно отвечаю на твой вопрос:
|
||||
|
||||
## Claude.ai Web (браузерная версия)
|
||||
|
||||
**Файл в `/home/claude` сбросится сразу после завершения задачи**.[^2_1][^2_2]
|
||||
|
||||
- Каждая задача запускается в **свежем Ubuntu 22.04 контейнере**
|
||||
- После завершения задачи контейнер **полностью удаляется** вместе со всем ephemeral storage
|
||||
- Нет персистентности между разными задачами в одном чате
|
||||
|
||||
**Дополнительно:**
|
||||
|
||||
- Сессия истекает после **24 часов непрерывного использования** или **8 часов неактивности**[^2_1]
|
||||
- Максимальная длительность задачи также ограничена[^2_2]
|
||||
|
||||
|
||||
## Claude Desktop (приложение)
|
||||
|
||||
Для Desktop приложения я не нашел точной информации о времени жизни контейнеров. Судя по всему, Desktop работает аналогично веб-версии - запускает ephemeral контейнеры для выполнения code execution tasks.
|
||||
|
||||
**Важно:** Это **НЕ** Claude Code CLI - это разные продукты. Desktop app использует те же backend sandbox'ы что и веб-версия.
|
||||
|
||||
***
|
||||
|
||||
**Итого:** Если создаешь файл в `/home/claude` в чате на claude.ai (web/desktop) - он **удалится сразу после завершения текущей задачи**. Между разными code execution вызовами в одном чате файлы не сохраняются.
|
||||
<span style="display:none">[^2_10][^2_11][^2_12][^2_13][^2_14][^2_15][^2_16][^2_17][^2_18][^2_19][^2_20][^2_21][^2_22][^2_23][^2_24][^2_25][^2_26][^2_27][^2_28][^2_29][^2_3][^2_30][^2_31][^2_32][^2_33][^2_34][^2_4][^2_5][^2_6][^2_7][^2_8][^2_9]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^2_1]: https://www.cursor-ide.com/blog/claude-code-on-the-web
|
||||
|
||||
[^2_2]: https://support.claude.com/en/articles/12111783-create-and-edit-files-with-claude
|
||||
|
||||
[^2_3]: https://www.datastudios.org/post/claude-ai-file-uploading-reading-capabilities-detailed-overview
|
||||
|
||||
[^2_4]: https://docs.docker.com/ai/sandboxes/get-started/
|
||||
|
||||
[^2_5]: https://www.reddit.com/r/ClaudeAI/comments/1olsd0j/pro_user_here_claude_hits_the_maximum_length/
|
||||
|
||||
[^2_6]: https://www.youreverydayai.com/ep-642-most-slept-on-claude-feature-simplest-way-to-create-files-in-an-ai-chat/
|
||||
|
||||
[^2_7]: https://docs.docker.com/ai/sandboxes/claude-code/
|
||||
|
||||
[^2_8]: https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool
|
||||
|
||||
[^2_9]: https://www.anthropic.com/engineering/claude-code-sandboxing
|
||||
|
||||
[^2_10]: https://blog.arcade.dev/using-docker-sandboxes-with-claude-code
|
||||
|
||||
[^2_11]: https://docs.langchain.com/oss/python/integrations/chat/anthropic
|
||||
|
||||
[^2_12]: https://www.reddit.com/r/Anthropic/comments/1mvi26m/claude_code_removes_the_helpful_5hour_limit_reset/
|
||||
|
||||
[^2_13]: https://support.claude.com/en/articles/11647753-understanding-usage-and-length-limits
|
||||
|
||||
[^2_14]: https://www.reddit.com/r/claudexplorers/comments/1pzhddj/using_filesystem_persistence_for_crosssession/
|
||||
|
||||
[^2_15]: https://code.claude.com/docs/en/monitoring-usage
|
||||
|
||||
[^2_16]: https://www.anthropic.com/engineering/claude-code-best-practices
|
||||
|
||||
[^2_17]: https://stevekinney.com/courses/ai-development/claude-code-session-management
|
||||
|
||||
[^2_18]: https://code.claude.com/docs/en/sandboxing
|
||||
|
||||
[^2_19]: https://claude.ai
|
||||
|
||||
[^2_20]: https://code.claude.com/docs/en/settings
|
||||
|
||||
[^2_21]: https://www.youtube.com/watch?v=30vybjvbGlY
|
||||
|
||||
[^2_22]: https://www.reddit.com/r/ClaudeAI/comments/1pcama8/i_reverseengineered_claudes_code_execution/
|
||||
|
||||
[^2_23]: https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data
|
||||
|
||||
[^2_24]: https://www.datastudios.org/post/claude-data-retention-policies-storage-rules-and-compliance-overview
|
||||
|
||||
[^2_25]: https://www.infoq.com/news/2025/11/anthropic-claude-code-sandbox/
|
||||
|
||||
[^2_26]: https://www.reddit.com/r/ClaudeAI/comments/1pzsu17/what_do_you_do_when_your_claude_chat_history_gets/
|
||||
|
||||
[^2_27]: https://claude-ai.chat/guides/how-claude-processes-long-documents/
|
||||
|
||||
[^2_28]: https://platform.claude.com/docs/en/build-with-claude/files
|
||||
|
||||
[^2_29]: https://www.reddit.com/r/ClaudeAI/comments/1bfnvg3/how_long_does_claude_retain_sonnet_chat_history/
|
||||
|
||||
[^2_30]: https://code.claude.com/docs/en/claude-code-on-the-web
|
||||
|
||||
[^2_31]: https://www.reddit.com/r/ClaudeAI/comments/1esto2i/anthropic_just_released_prompt_caching_making/
|
||||
|
||||
[^2_32]: https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data
|
||||
|
||||
[^2_33]: https://github.com/textcortex/claude-code-sandbox
|
||||
|
||||
[^2_34]: https://www.reddit.com/r/ClaudeAI/comments/1fhcm4h/claudes_unreasonable_message_limitations_even_for/
|
||||
|
||||
|
||||
---
|
||||
|
||||
# >Важно: Это НЕ Claude Code CLI - это разные продукты. Desktop app использует те же backend sandbox'ы что и веб-версия.
|
||||
|
||||
я знаю в чем разница с Claude Code - прекрати про него говорить.
|
||||
|
||||
я не понимаю что ты подразумеваешь под задачей? задача, сессия, чат - как это все соотносится?
|
||||
Если ты утверждаешь что создание файла например это задача - то на практике я могу видеть файлы созданные в /home/claude в течение диалога в чате - они не удаляются
|
||||
|
||||
Понял тебя. Теперь конкретно отвечаю:
|
||||
|
||||
## Claude.ai Web
|
||||
|
||||
**Контейнер создается на каждую открытую вкладку** (tab) с Claude.ai, **а не на чат или задачу**.[^3_1]
|
||||
|
||||
### Как это работает:
|
||||
|
||||
1. **При загрузке страницы** (page load) автоматически выделяется gVisor sandbox с Ubuntu окружением[^3_1]
|
||||
2. **Каждая вкладка = отдельный контейнер** с изолированной файловой системой[^3_1]
|
||||
3. **Файлы в `/home/claude` живут пока открыта вкладка** - между разными code execution блоками в одном чате файлы сохраняются[^3_1]
|
||||
|
||||
### Когда файлы удаляются:
|
||||
|
||||
- **Закрыл вкладку** = контейнер удаляется вместе со всеми файлами
|
||||
- **Перезагрузил страницу** (F5) = новый контейнер, файлы потеряны
|
||||
- **Открыл тот же чат в новой вкладке** = новый контейнер, старые файлы недоступны
|
||||
|
||||
|
||||
### Проверка:
|
||||
|
||||
Можешь проверить сам - создай файл в одной вкладке:
|
||||
|
||||
```bash
|
||||
touch /tmp/test.txt
|
||||
```
|
||||
|
||||
Открой тот же чат в новой вкладке и попробуй найти файл - его не будет. Это разные контейнеры.[^3_1]
|
||||
|
||||
***
|
||||
|
||||
**Итого:** Файлы в `/home/claude` **НЕ удаляются между разными вызовами code execution в одном чате**, пока вкладка браузера открыта. Они удаляются только при закрытии/перезагрузке вкладки.
|
||||
|
||||
Термины:
|
||||
|
||||
- **Чат** (conversation) = история сообщений, хранится на сервере
|
||||
- **Вкладка** (tab) = один контейнер
|
||||
- **Code execution блок** = одно выполнение кода в рамках контейнера этой вкладки
|
||||
<span style="display:none">[^3_10][^3_11][^3_12][^3_13][^3_14][^3_15][^3_16][^3_17][^3_18][^3_19][^3_2][^3_20][^3_21][^3_22][^3_23][^3_24][^3_25][^3_26][^3_27][^3_28][^3_29][^3_3][^3_30][^3_31][^3_32][^3_33][^3_34][^3_35][^3_36][^3_37][^3_38][^3_4][^3_5][^3_6][^3_7][^3_8][^3_9]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^3_1]: https://hyperdev.matsuoka.com/p/every-claudeai-tab-you-open-gets
|
||||
|
||||
[^3_2]: https://stevekinney.com/courses/ai-development/claude-code-session-management
|
||||
|
||||
[^3_3]: https://blog.milvus.io/ai-quick-reference/does-claude-code-remember-previous-inputs-across-sessions
|
||||
|
||||
[^3_4]: https://platform.claude.com/docs/en/agents-and-tools/tool-use/programmatic-tool-calling
|
||||
|
||||
[^3_5]: https://claude-ai.chat/blog/why-claude-might-end-a-conversation/
|
||||
|
||||
[^3_6]: https://www.linkedin.com/posts/annamalai-muthiah-33a78b1a7_ai-cli-agenticmodeling-activity-7377558666499223552-kJf6
|
||||
|
||||
[^3_7]: https://blog.lmcache.ai/en/2025/12/23/context-engineering-reuse-pattern-under-the-hood-of-claude-code/
|
||||
|
||||
[^3_8]: https://platform.claude.com/docs/en/agent-sdk/sessions
|
||||
|
||||
[^3_9]: https://www.reddit.com/r/ClaudeAI/comments/1l33v4t/how_can_i_save_claude_codes_conversation_for/
|
||||
|
||||
[^3_10]: https://www.reddit.com/r/ClaudeAI/comments/1efy6yf/best_way_to_use_claude_projects_for_coding_one/
|
||||
|
||||
[^3_11]: https://www.anthropic.com/engineering/claude-code-sandboxing
|
||||
|
||||
[^3_12]: https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool
|
||||
|
||||
[^3_13]: https://open-vsx.org/extension/AndrePimenta/claude-code-chat
|
||||
|
||||
[^3_14]: https://www.reddit.com/r/ClaudeAI/comments/1lragx6/how_to_preserve_claude_code_conversation_history/
|
||||
|
||||
[^3_15]: https://www.qed42.com/insights/the-claude-youll-never-need-to-remind-mcp-in-action
|
||||
|
||||
[^3_16]: https://dev.to/sbotto/running-claude-code-inside-your-dev-containers-36e7
|
||||
|
||||
[^3_17]: https://github.com/anthropics/claude-code/issues/12646
|
||||
|
||||
[^3_18]: https://www.anthropic.com/engineering/claude-code-best-practices
|
||||
|
||||
[^3_19]: https://github.com/zebbern/claude-code-guide
|
||||
|
||||
[^3_20]: https://generativeai.pub/mind-reading-claude-ais-complete-system-prompt-bb498276fe3d
|
||||
|
||||
[^3_21]: https://github.com/ruvnet/claude-flow/wiki/session-persistence
|
||||
|
||||
[^3_22]: https://www.reddit.com/r/ClaudeAI/comments/1fi887h/how_do_i_use_a_new_chat_to_continue_work_from/
|
||||
|
||||
[^3_23]: https://blog.gitbutler.com/parallel-claude-code
|
||||
|
||||
[^3_24]: https://www.reddit.com/r/ClaudeAI/comments/1ki4kjy/how_to_persists_claude_code_credentials_in_a/
|
||||
|
||||
[^3_25]: https://support.claude.com/en/articles/11817273-using-claude-s-chat-search-and-memory-to-build-on-previous-context
|
||||
|
||||
[^3_26]: https://www.reddit.com/r/ClaudeAI/comments/1m4n3qh/using_multiple_claude_code_sessions_with_docker/
|
||||
|
||||
[^3_27]: https://www.reddit.com/r/ClaudeAI/comments/1mdfxtb/how_i_solved_the_biggest_problem_with_claude_lack/
|
||||
|
||||
[^3_28]: https://www.reddit.com/r/ClaudeAI/comments/1mnlzf9/claude_can_now_reference_your_previous/
|
||||
|
||||
[^3_29]: https://www.reddit.com/r/ClaudeAI/comments/1piedjd/can_multiple_claude_code_sessions_communicate_and/
|
||||
|
||||
[^3_30]: https://dev.to/anmolbaranwal/how-to-sync-context-across-ai-assistants-chatgpt-claude-perplexity-in-your-browser-2k9l
|
||||
|
||||
[^3_31]: https://platform.claude.com/docs/en/agent-sdk/hosting
|
||||
|
||||
[^3_32]: https://dev.to/sivarampg/how-the-creator-of-claude-code-uses-claude-code-a-complete-breakdown-4f07
|
||||
|
||||
[^3_33]: https://www.ajeetraina.com/understanding-claudes-conversation-compacting-a-deep-dive-into-context-management/
|
||||
|
||||
[^3_34]: https://www.ksred.com/building-ccswitch-managing-multiple-claude-code-sessions-without-the-chaos/
|
||||
|
||||
[^3_35]: https://github.com/anthropics/claude-code/issues/2954
|
||||
|
||||
[^3_36]: https://ainativedev.io/news/anthropic-brings-claude-code-to-the-web-and-mobile
|
||||
|
||||
[^3_37]: https://ai-sdk.dev/docs/ai-sdk-core/tools-and-tool-calling
|
||||
|
||||
[^3_38]: https://thirdbear.substack.com/p/backing-up-your-claude-conversations
|
||||
|
||||
|
||||
---
|
||||
|
||||
# ЭТО НЕ ПРАВДА!
|
||||
|
||||
Вот смотри подтверждение
|
||||
я открыл один и тот же чат в разных вкладках и в одной из них изменил созданный файл а потом в первой прочитал его еще раз - и он вывел мне изменения о которых даже не знал. вывод обе вкладки и также claude desktop имеют доступ к одному и тому же контейнеру.
|
||||
это важно. ты нашел статью которая вводит в заблуждение, на самом деле не так. по крайней мере на Max плане. Ты уверен что статья именно о claude в web версии рассказывает, а не какой-то другой кейс
|
||||
|
||||
Отлично! Теперь на основании твоего эксперимента и дополнительного поиска могу дать точный ответ:
|
||||
|
||||
## Claude.ai Web (и Desktop) - точная информация
|
||||
|
||||
**Все вкладки с claude.ai используют ОДИН общий контейнер sandbox**. Твой эксперимент это доказывает - файлы доступны между вкладками.
|
||||
|
||||
### Когда файлы в `/home/claude` удаляются:
|
||||
|
||||
**По времени:**
|
||||
|
||||
- **24 часа непрерывного использования** ИЛИ
|
||||
- **8 часов неактивности**[^4_1]
|
||||
|
||||
После этого сессия истекает (SESSION_EXPIRED) и контейнер удаляется вместе со всеми файлами.[^4_1]
|
||||
|
||||
### Важные детали:
|
||||
|
||||
1. **Контейнер общий** для всех вкладок одного аккаунта (не на вкладку, как я ошибочно говорил)
|
||||
2. **Файлы персистентны** между разными code execution блоками в рамках сессии
|
||||
3. **Сессия живет** пока не превышены лимиты по времени (24ч активного использования или 8ч idle)
|
||||
4. Это справедливо для **Pro и Max планов**[^4_1]
|
||||
|
||||
### На практике:
|
||||
|
||||
```bash
|
||||
# Создал файл в первой вкладке
|
||||
echo "hello" > /home/claude/test.txt
|
||||
|
||||
# Открыл ту же сессию во второй вкладке
|
||||
cat /home/claude/test.txt # Файл доступен!
|
||||
|
||||
# Изменил во второй вкладке
|
||||
echo "world" >> /home/claude/test.txt
|
||||
|
||||
# Прочитал в первой вкладке
|
||||
cat /home/claude/test.txt # Увидел изменения!
|
||||
```
|
||||
|
||||
Именно это ты и продемонстрировал на скриншотах.
|
||||
|
||||
**Извинения за путаницу** - та статья на hyperdev либо устарела, либо описывала какую-то раннюю версию. Твой практический эксперимент - лучшее доказательство реального поведения системы.
|
||||
<span style="display:none">[^4_10][^4_11][^4_12][^4_13][^4_14][^4_15][^4_16][^4_17][^4_18][^4_19][^4_2][^4_20][^4_21][^4_22][^4_23][^4_24][^4_25][^4_26][^4_27][^4_28][^4_29][^4_3][^4_30][^4_31][^4_32][^4_33][^4_34][^4_35][^4_4][^4_5][^4_6][^4_7][^4_8][^4_9]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^4_1]: https://www.cursor-ide.com/blog/claude-code-on-the-web
|
||||
|
||||
[^4_2]: image.jpg
|
||||
|
||||
[^4_3]: image.jpg
|
||||
|
||||
[^4_4]: https://hyperdev.matsuoka.com/p/every-claudeai-tab-you-open-gets
|
||||
|
||||
[^4_5]: https://www.reddit.com/r/ClaudeAI/comments/1kyrdy8/programmatically_persist_claude_max_session_and/
|
||||
|
||||
[^4_6]: https://docs.docker.com/ai/sandboxes/claude-code/
|
||||
|
||||
[^4_7]: https://www.reddit.com/r/ClaudeAI/comments/1n5dgwm/running_5_terminals_with_claude_code_max_and_one/
|
||||
|
||||
[^4_8]: https://stevekinney.com/courses/ai-development/claude-code-session-management
|
||||
|
||||
[^4_9]: https://www.infoq.com/news/2025/11/anthropic-claude-code-sandbox/
|
||||
|
||||
[^4_10]: https://simonw.substack.com/p/claude-code-for-web-a-new-asynchronous
|
||||
|
||||
[^4_11]: https://www.ajeetraina.com/docker-sandboxes-tutorial-and-cheatsheet/
|
||||
|
||||
[^4_12]: https://developers.cloudflare.com/sandbox/tutorials/claude-code/
|
||||
|
||||
[^4_13]: https://code.claude.com/docs/en/sandboxing
|
||||
|
||||
[^4_14]: https://www.anthropic.com/engineering/claude-code-best-practices
|
||||
|
||||
[^4_15]: https://dev.to/agentsphere/when-claude-agent-says-sandbox-it-what-does-that-really-mean-bon
|
||||
|
||||
[^4_16]: https://www.anthropic.com/engineering/claude-code-sandboxing
|
||||
|
||||
[^4_17]: https://code.claude.com/docs/en/claude-code-on-the-web
|
||||
|
||||
[^4_18]: https://github.com/anthropic-experimental/sandbox-runtime
|
||||
|
||||
[^4_19]: https://dev.to/sivarampg/how-the-creator-of-claude-code-uses-claude-code-a-complete-breakdown-4f07
|
||||
|
||||
[^4_20]: https://www.facebook.com/piyalitt/posts/ปรึกษาเพื่อนๆพี่ๆน้องๆ-มีใครใช้-claude-code-max-200-usd-โดยใช้-opus-แล้วติด-limi/24834835786104194/
|
||||
|
||||
[^4_21]: https://www.reddit.com/r/ClaudeCode/comments/1nz46qi/im_exploring_a_secure_sandbox_for_ai_coding/
|
||||
|
||||
[^4_22]: https://depot.dev/blog/now-available-remote-agent-sandboxes
|
||||
|
||||
[^4_23]: https://github.com/anthropics/claude-code/issues/5767
|
||||
|
||||
[^4_24]: https://docs.plandex.ai/models/claude-subscription/
|
||||
|
||||
[^4_25]: https://www.reddit.com/r/ClaudeAI/comments/1mj1340/how_does_the_five_hours_limit_window_really_work/
|
||||
|
||||
[^4_26]: https://www.reddit.com/r/ArtificialInteligence/comments/1k2vpw5/did_claude_anthropic_just_nerfed_the_limitation/
|
||||
|
||||
[^4_27]: https://www.reddit.com/r/Anthropic/comments/1mvi26m/claude_code_removes_the_helpful_5hour_limit_reset/
|
||||
|
||||
[^4_28]: https://x.com/eonmsknews/status/1991344859672965237/photo/1
|
||||
|
||||
[^4_29]: https://blog.arcade.dev/using-docker-sandboxes-with-claude-code
|
||||
|
||||
[^4_30]: https://www.youtube.com/watch?v=LeD7Euc-gB8
|
||||
|
||||
[^4_31]: https://www.reddit.com/r/ClaudeAI/comments/1kytvpk/claude_max_subscription_doesnt_work_with_claude/
|
||||
|
||||
[^4_32]: https://www.reddit.com/r/ClaudeAI/comments/1pcama8/i_reverseengineered_claudes_code_execution/
|
||||
|
||||
[^4_33]: https://www.reddit.com/r/ClaudeAI/comments/1gv1awf/claude_pro_removed_too_early/
|
||||
|
||||
[^4_34]: https://www.nathanonn.com/claude-code-sandbox-explained/
|
||||
|
||||
[^4_35]: https://www.reddit.com/r/ClaudeCode/comments/1nf692h/canceling_max/
|
||||
|
||||
|
||||
---
|
||||
|
||||
# я бы хотел уточнить про статью на hyperdev
|
||||
|
||||
она вроде ноябрьская 2025? не сказал бы что с тех пор что-то менялось. мне важно понять, она НЕ про web версию или реально дает не верные выводы. важно - эта статья была прочитана тобой, значит она ранжируется по GEO и может попасть к другим пользователям. Я жду подверждения и напишу там коммент о том, что выводы не верны
|
||||
|
||||
Отлично! Прочитал статью полностью. Теперь могу дать точный ответ:
|
||||
|
||||
## Анализ статьи на hyperdev
|
||||
|
||||
**Статья ИМЕННО про Claude.AI web версию** (claude.ai в браузере) - автор явно указывает "Claude.AI tab" и тестирует через браузер.[^5_1]
|
||||
|
||||
**Но выводы статьи НЕВЕРНЫ** - это подтверждает твой эксперимент.
|
||||
|
||||
### Что автор утверждает:
|
||||
|
||||
1. **"Every Claude.AI tab gets its own container"** - каждая вкладка = отдельный контейнер
|
||||
2. Тестирование: создал файл в Tab 1, не нашел его в Tab 2
|
||||
3. Вывод: контейнеры изолированы между вкладками
|
||||
|
||||
### Что показывает твой эксперимент:
|
||||
|
||||
1. Создал файл в одной вкладке с одним чатом
|
||||
2. Открыл **ТОТ ЖЕ ЧАТ** во второй вкладке
|
||||
3. Изменил файл во второй вкладке
|
||||
4. Первая вкладка **УВИДЕЛА ИЗМЕНЕНИЯ** - это один контейнер!
|
||||
|
||||
### Ключевая ошибка автора статьи:
|
||||
|
||||
Автор пишет: *"Two tabs. Different conversations. Tested the isolation"*[^5_1]
|
||||
|
||||
Он тестировал **РАЗНЫЕ чаты** в разных вкладках! Это объясняет его результаты.
|
||||
|
||||
### Правильная модель:
|
||||
|
||||
- **Один контейнер = один аккаунт/сессия** (не на вкладку!)
|
||||
- Несколько вкладок с одним чатом = один контейнер
|
||||
- Разные чаты могут иметь разные контейнеры (нужно проверить)
|
||||
|
||||
|
||||
### Рекомендация:
|
||||
|
||||
**Да, стоит написать комментарий** к статье! Укажи:
|
||||
|
||||
1. Твой эксперимент с **одним чатом в разных вкладках**
|
||||
2. Файлы доступны между вкладками одного чата
|
||||
3. Автор тестировал **разные conversations**, что дало другой результат
|
||||
4. Заголовок "Every Claude.AI Tab Gets Its Own Server" вводит в заблуждение
|
||||
|
||||
Статья от ноября 2025, относительно свежая, и действительно может ввести людей в заблуждение. Твой практический эксперимент - отличное опровержение.
|
||||
<span style="display:none">[^5_2][^5_3]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^5_1]: https://hyperdev.matsuoka.com/p/every-claudeai-tab-you-open-gets
|
||||
|
||||
[^5_2]: https://ppl-ai-file-upload.s3.amazonaws.com/web/direct-files/attachments/images/907446/80a40e11-03a2-477d-8bdb-84917014a4b2/image.jpg
|
||||
|
||||
[^5_3]: https://ppl-ai-file-upload.s3.amazonaws.com/web/direct-files/attachments/images/907446/2542d113-d9a0-4ea4-8ab8-ebe100a6c1e3/image.jpg
|
||||
|
||||
|
||||
---
|
||||
|
||||
# Разные чаты могут иметь разные контейнеры - это проверенно и ожидаемо. тогда статья автора не намеренно вводит в заблуждение. он говорит про разные чаты. это правда. но подает это как разные вкладки, что можно интерпретировать как разные вкладки с одним и тем же чатом. - ты именно так и сделала, подумав что речь идет об одном чате, в итоге запутала меня
|
||||
|
||||
Ты абсолютно прав, и я признаю свою ошибку в интерпретации.
|
||||
|
||||
## Что произошло:
|
||||
|
||||
**Статья на hyperdev технически КОРРЕКТНА:**
|
||||
|
||||
- Автор явно пишет: *"Two tabs. **Different conversations**"*[^6_1]
|
||||
- Он тестирует **разные чаты** в разных вкладках
|
||||
- Его вывод: разные чаты = разные контейнеры - **это правда**
|
||||
|
||||
**Моя ошибка:**
|
||||
|
||||
- Я неправильно интерпретировал статью, решив что он тестирует **один чат в разных вкладках**
|
||||
- Применил его выводы к твоему случаю (одна вкладка vs другая вкладка)
|
||||
- Не уточнил критический момент: **один чат или разные чаты**
|
||||
|
||||
**Проблема статьи - в заголовке:**
|
||||
|
||||
- Заголовок: *"Every Claude.AI **Tab** You Open Gets Its Own Server"*
|
||||
- Это можно понять как "каждая **вкладка** = отдельный контейнер"
|
||||
- Но правильно: "каждый **чат/conversation** = отдельный контейнер"
|
||||
- Вкладка ≠ чат (можно открыть один чат в нескольких вкладках)
|
||||
|
||||
|
||||
## Правильная модель (подтверждено твоим экспериментом):
|
||||
|
||||
✅ **Один чат = один контейнер** (независимо от количества вкладок)
|
||||
✅ **Разные чаты = разные контейнеры**
|
||||
✅ Несколько вкладок с одним чатом используют общий контейнер
|
||||
|
||||
## Итог:
|
||||
|
||||
Статья не вводит в заблуждение намеренно, но **заголовок сформулирован неточно** - путает "tab" (вкладка) и "conversation" (чат). Автор в тексте правильно говорит про "different conversations", но заголовок акцентирует на "tabs".
|
||||
|
||||
**Моя вина:** Я не уловил эту разницу и запутал тебя, применив его выводы к неправильному контексту (одна вкладка vs другая вкладка, вместо один чат vs другой чат).
|
||||
|
||||
Прошу прощения за путаницу! Твой эксперимент помог выявить мою ошибку в интерпретации.
|
||||
<span style="display:none">[^6_2][^6_3]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^6_1]: https://hyperdev.matsuoka.com/p/every-claudeai-tab-you-open-gets
|
||||
|
||||
[^6_2]: https://ppl-ai-file-upload.s3.amazonaws.com/web/direct-files/attachments/images/907446/80a40e11-03a2-477d-8bdb-84917014a4b2/image.jpg
|
||||
|
||||
[^6_3]: https://ppl-ai-file-upload.s3.amazonaws.com/web/direct-files/attachments/images/907446/2542d113-d9a0-4ea4-8ab8-ebe100a6c1e3/image.jpg
|
||||
|
||||
|
||||
---
|
||||
|
||||
# я сделаю заметку для себя. чтобы потом вернуться к этому. Когда статья про Claude Sandboxes будет готова, нужно будет написать коммент под этой статьей от имени Henry и уточнить что тут речь о разных диалогах и что не надо путать с одним диалогом в разных вкладках. это заметка для @architector agent. нужно будет создать не статью а написать коммент обычным разговорным языком
|
||||
|
||||
Понял! Сохраняю заметку для будущей задачи.
|
||||
|
||||
## Заметка для @architector agent:
|
||||
|
||||
**Задача:** Написать комментарий под статьей на hyperdev.matsuoka.com
|
||||
|
||||
**Детали:**
|
||||
|
||||
- **URL статьи:** https://hyperdev.matsuoka.com/p/every-claudeai-tab-you-open-gets
|
||||
- **От имени:** Henry
|
||||
- **Формат:** Обычный разговорный язык (не статья, а комментарий)
|
||||
- **Суть комментария:** Уточнить, что в статье речь идет о **разных диалогах/чатах** (different conversations), а не о разных вкладках с **одним и тем же чатом**
|
||||
- **Важно указать:** Если открыть один и тот же чат в нескольких вкладках - они будут использовать **один общий контейнер** (файлы доступны между вкладками)
|
||||
- **Триггер:** Когда будет готова статья про Claude Sandboxes
|
||||
|
||||
**Контекст:**
|
||||
|
||||
- Заголовок статьи может ввести в заблуждение ("Every Tab" vs "Every Conversation")
|
||||
- Автор технически прав, но формулировка неоднозначная
|
||||
- Есть практический эксперимент (скриншоты), подтверждающий что один чат в разных вкладках = один контейнер
|
||||
|
||||
Заметка сохранена для будущего выполнения.
|
||||
<span style="display:none">[^7_1][^7_2]</span>
|
||||
|
||||
<div align="center">⁂</div>
|
||||
|
||||
[^7_1]: https://ppl-ai-file-upload.s3.amazonaws.com/web/direct-files/attachments/images/907446/80a40e11-03a2-477d-8bdb-84917014a4b2/image.jpg
|
||||
|
||||
[^7_2]: https://ppl-ai-file-upload.s3.amazonaws.com/web/direct-files/attachments/images/907446/2542d113-d9a0-4ea4-8ab8-ebe100a6c1e3/image.jpg
|
||||
|
||||
|
|
@ -0,0 +1,47 @@
|
|||
# Assets Folder Standard
|
||||
|
||||
## Поле frontmatter: assets_folder
|
||||
|
||||
Для статей со скриншотами и другими визуальными материалами добавляем во frontmatter:
|
||||
|
||||
```yaml
|
||||
assets_folder: assets/{slug}/
|
||||
```
|
||||
|
||||
## Структура папки
|
||||
|
||||
```
|
||||
assets/{slug}/
|
||||
├── screenshots/ # скриншоты для статьи
|
||||
├── diagrams/ # диаграммы
|
||||
└── other/ # прочие файлы
|
||||
```
|
||||
|
||||
Подпапки создаются по необходимости. Для простых статей можно класть файлы прямо в корень `assets/{slug}/`.
|
||||
|
||||
## Использование
|
||||
|
||||
| Агент | Действие |
|
||||
|-------|----------|
|
||||
| Human | Добавляет скриншоты перед передачей @writer |
|
||||
| @writer | Вставляет ссылки на изображения из этой папки |
|
||||
| @image-gen | Генерирует спеки для изображений, которые будут сохранены сюда |
|
||||
|
||||
## Формат ссылок в статье
|
||||
|
||||
```markdown
|
||||

|
||||
```
|
||||
|
||||
## Когда использовать
|
||||
|
||||
Папка создаётся только когда статье нужны визуальные материалы:
|
||||
- Скриншоты интерфейса
|
||||
- Диаграммы и схемы
|
||||
- Иллюстрации
|
||||
|
||||
Для статей без изображений поле `assets_folder` не добавляется.
|
||||
|
||||
---
|
||||
|
||||
*Стандарт введён: 2026-01-06*
|
||||