Claude Cowork is an autonomous desktop agent from Anthropic, released January 2026 as a research preview. It extends the capabilities of Claude Code — previously limited to developers working in the terminal — to general knowledge workers through a visual interface in the Claude Desktop app.
How It Works
Users grant Claude access to specific folders on their computer and describe tasks in natural language. Rather than providing instructions or suggestions, Cowork executes the work directly: reading files, creating plans, breaking complex tasks into parallel subtasks, and delivering finished outputs.
Typical tasks include file organization, document generation (Excel spreadsheets with formulas, PowerPoint presentations, Word documents), data extraction from PDFs and images, and synthesis of research across multiple files.
Security Architecture
Cowork runs inside an isolated virtual machine using Apple's Virtualization Framework with a custom Linux filesystem. This architectural separation means Claude can only access folders explicitly authorized by the user — the rest of the system is inaccessible at the VM level, not just by policy. Users can grant or revoke folder access at any time, and file processing occurs locally.
Availability and Limitations
- Requires Claude Max subscription ($100–200/month)
- macOS only; Windows support planned
- No session memory — each task starts fresh
- No integration with Claude Projects
- External connectors still maturing
Skills/Hooks Support
Cowork shares configuration with Claude Code. They use the same ~/.claude/settings.json and skills from ~/.claude/skills/.
Key Differences
Configuration Sharing
Claude Cowork shares configuration with Claude Code. It directly reuses ~/.claude/settings.json. This means same API keys, same base URL, same models, same behavior. Claude Cowork also auto-discovers skills from ~/.claude/skills/. GitHub
What Cowork Inherits
- Skills (auto-discovered from
~/.claude/skills/) - MCP connectors
- Settings.json configuration
- Hooks (at least those defined in settings.json)
What's Different
- Architecture: Cowork uses VZVirtualMachine (Apple Virtualization Framework) and downloads/boots a custom Linux root filesystem. It's regular Claude Code wrapped in a less intimidating interface with a filesystem sandbox configured automatically.
- Model Default: Cowork defaults to Opus 4.5 for better autonomous task handling, while Claude Code defaults to Sonnet 4.5.
- Prompts/System Instructions: Different system prompts oriented toward file management vs coding tasks. Skills load automatically into Cowork, and as models get smarter, the same skill files produce better results without changes.
- No Memory/Projects: Claude does not retain context between Cowork sessions. Every task starts from a blank slate. You cannot currently use Cowork within Claude Projects.
Reuse of hooks
You can now use hooks inside skills prompts and subagents, which is extremely powerful.