c2de35caba
- Create comprehensive Cowork docs (23 files in cowork/) - 6 guides: overview, getting started, capabilities, security, troubleshooting - 60+ ready-to-use prompts across 4 categories - 5 step-by-step workflows - Reference materials: cheatsheet, FAQ, comparison, glossary - Integrate Perplexity research (P0/P1/P2) - Exact error messages with solutions (VPN, Chrome host, context limits) - Competitive analysis (vs Copilot/Gemini/ChatGPT/Apple Intelligence) - Enterprise validation (TELUS, Rakuten, Zapier stats) - OCR accuracy benchmarks (97% field, 63% line-item) - Token budget planning per task type - Document critical limitations - VPN incompatibility (#1 community issue) - Context limit reality (165K vs 200K theoretical) - Platform constraints (macOS only) - Usage limits and pricing (Pro $20, Max $100-200) - Update central files - README.md: detailed Cowork section with tables - VERSION: 3.9.6 → 3.9.7 - machine-readable/reference.yaml: add cowork_reference entry - machine-readable/cowork-reference.yaml: new LLM-optimized index (~1.5K tokens) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
9.7 KiB
9.7 KiB
Claude Cowork: Overview
Reading time: ~5 minutes
What is Cowork?
Cowork is Claude's agentic desktop feature that extends autonomous AI capabilities to non-technical users through the Claude Desktop app. Instead of terminal commands, Cowork accesses local folders and files directly.
Key Facts
| Aspect | Details |
|---|---|
| Release | Research preview, January 2026 |
| Availability | Pro ($20/mo) or Max ($100-200/mo) subscribers |
| Platform | macOS only (Windows planned, Linux not announced) |
| Built by | ~90% written by Claude itself ([Anthropic blog]) |
| Focus | File manipulation, organization, document generation |
| Usage limits | Resets every 5 hours; heavy tasks consume quota rapidly |
Relationship to Claude Code
Cowork shares the same backend architecture as Claude Code:
| Shared | Different |
|---|---|
| Same model capabilities | Desktop app vs Terminal |
| Extended thinking | Files only vs full shell |
| Agentic loops | Knowledge workers vs developers |
| Sub-agent architecture | No code execution |
Key difference: Cowork cannot execute arbitrary code—it manipulates files only.
Who Should Use Cowork?
Good Fit
| Persona | Use Case | Why Cowork |
|---|---|---|
| Project Manager | File organization + status reports | Multi-step automation without coding |
| Data Analyst | Local CSV/Excel → formatted reports | Native Excel formulas output |
| Writer/Editor | Research notes → structured documents | Synthesis across many sources |
| Operations | Receipt screenshots → expense reports | Multi-format input processing |
| Consultant | Client docs → deliverables | Cross-reference and synthesis |
| Researcher | Papers + notes → literature review | Source organization and citation |
Poor Fit (For Now)
| Persona | Limitation | Alternative |
|---|---|---|
| Security-conscious enterprises | No audit trail, no access controls | Wait for enterprise features |
| Heavy cloud users | No Google Drive/Dropbox confirmed | Use native cloud apps |
| Anyone needing reliability | Preview = bugs, unexpected behavior | Wait for stable release |
| Code execution needs | Files only, no scripts | Use Claude Code |
| Windows/Linux users | macOS only (Windows planned, Linux not announced) | Wait for platform expansion |
| Heavy daily users | Usage limits reset every 5h; Pro exhausts in ~1-1.5h intensive use | Consider Max tier or batch work |
| VPN users | Cannot work with VPN active (VM routing conflict) | Disconnect VPN or use Claude Code |
Architecture
Local-First Design
┌─────────────────────────────────────────────────────────────┐
│ CLAUDE DESKTOP APP │
│ ┌─────────────────────────────────────────────────────────┐│
│ │ COWORK ││
│ │ ┌─────────────────────────────────────────────────────┐││
│ │ │ ORCHESTRATOR │││
│ │ │ • Receives user request │││
│ │ │ • Creates execution plan │││
│ │ │ • Coordinates sub-agents │││
│ │ └──────────────────────┬──────────────────────────────┘││
│ │ ↓ ││
│ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────────┐ ││
│ │ │ Sub-Agent 1 │ │ Sub-Agent 2 │ │ Sub-Agent 3 │ ││
│ │ │ (Analysis) │ │ (Transform) │ │ (Organize) │ ││
│ │ │ │ │ │ │ │ ││
│ │ │ • Read files│ │ • Convert │ │ • Move files │ ││
│ │ │ • Extract │ │ • Format │ │ • Create dirs │ ││
│ │ │ • Summarize │ │ • Generate │ │ • Rename │ ││
│ │ └──────┬──────┘ └──────┬──────┘ └───────┬─────────┘ ││
│ │ └────────────────┴─────────────────┘ ││
│ │ ↓ ││
│ │ ┌─────────────────────────────────────────────────────┐││
│ │ │ LOCAL FOLDER SANDBOX │││
│ │ │ ~/Cowork-Workspace/ │││
│ │ │ ├── input/ (your files) │││
│ │ │ └── output/ (generated files) │││
│ │ └─────────────────────────────────────────────────────┘││
│ └─────────────────────────────────────────────────────────┘│
│ │
│ ┌─────────────────┐ │
│ │ CHROME INTEGRATION │ ← Web tasks (with explicit action) │
│ └─────────────────┘ │
└─────────────────────────────────────────────────────────────┘
Technical Specifications
| Aspect | Details |
|---|---|
| Access model | Local folder sandbox (user grants permission) |
| Cloud connectors | Not confirmed (no Drive, Dropbox, etc. yet) |
| Sub-agents | Parallel execution with fresh context per agent |
| Code execution | None — Files only (no scripts, no commands) |
| Browser | Chrome integration for web research tasks |
| Context limit | ~200K tokens (~150-500 pages per session) |
| Output formats | Excel (.xlsx with formulas), PowerPoint, Word, PDF, text, images |
How Sub-Agents Work
- Fresh context — Each sub-agent starts clean (no memory from other agents)
- Parallel execution — Multiple agents can work simultaneously
- Orchestrator coordination — Main agent assembles results
- Scope isolation — Each agent sees only what it needs
This architecture enables complex multi-step workflows while maintaining security boundaries.
What Cowork Cannot Do
Understanding limitations is crucial for effective use:
| Limitation | Implication |
|---|---|
| Execute shell commands | No mkdir, mv, cp via terminal |
| Run scripts | No Python, JavaScript, bash execution |
| Access cloud storage | No direct Google Drive, Dropbox, iCloud |
| Network requests | No API calls, no HTTP requests |
| Persist memory | No cross-session memory (only via files) |
| Access arbitrary folders | Only granted sandbox locations |
Workarounds
| Need | Workaround |
|---|---|
| Cloud files | Download to local workspace first |
| Code execution | Use Claude Code instead |
| Cross-session memory | Save context to a file, reload next session |
| Network data | Use Chrome integration for web research |
Mental Model
Think of Cowork as a highly capable assistant with physical access to one folder:
- Can read any file you put there
- Can create new files in any format
- Can reorganize, rename, transform
- Cannot leave that folder without permission
- Cannot run programs or scripts
- Cannot make network connections directly
This constraint is a feature, not a bug — it creates a safe sandbox for autonomous operation.
Enterprise Validation (Claude Adoption)
While Cowork is in research preview, Claude's underlying capabilities are validated at enterprise scale:
| Company | Results | Context |
|---|---|---|
| TELUS | $90M value, 500K hours saved | Document processing at scale |
| Rakuten | 87.5% reduction in processing time | Knowledge work automation |
| Zapier | 89% employee adoption | Workflow integration |
Relevance to Cowork: These stats are for Claude in general, not Cowork specifically. However, they validate the core AI capabilities that power Cowork's document processing and autonomous workflows.
What this means for you:
- The AI model works reliably at enterprise scale
- Document understanding and generation are production-ready
- The "research preview" label applies to the Cowork interface, not the underlying Claude capabilities
Next Steps
- Getting Started — Setup and first workflow
- Capabilities — Detailed feature breakdown
- Security — Safe usage practices