Open Cowork is a desktop application for Windows and macOS that puts a general-purpose AI agent behind a graphical interface, with prebuilt installers and no environment setup. It is an open-source implementation of Claude Cowork: you choose a workspace folder, describe a task in chat, and the agent reads and writes files, generates documents, and drives other applications inside that workspace.
The model is configured rather than fixed. Claude works through Anthropic or OpenRouter, any OpenAI-compatible endpoint can be pointed at by base URL, and presets exist for Zhipu GLM, MiniMax, and Kimi; Gemini is suggested where the agent controls a graphical interface. Calls go from the machine to the configured provider, and workspace files stay local.
Command execution is sandboxed in layers. On every platform, file operations are restricted to the workspace folder. Where WSL2 on Windows or Lima on macOS is available, Bash commands are routed into an isolated Linux virtual machine instead, with the workspace synced or mounted; without a virtual machine, commands fall back to native execution under the path restrictions.
Capabilities extend two ways. Skills are file-backed workflows for generating and processing PPTX, DOCX, PDF, and XLSX documents, and a skill-creator helps write new ones. MCP connectors attach external services such as a browser or Notion. Remote control through Feishu (Lark) and Slack sends commands and returns results from a chat platform, and a trace panel shows reasoning and tool execution as a task runs.
Features
- One-click installers: prebuilt Windows and macOS packages, plus a Homebrew cask, with no environment setup
- Flexible model support: Claude, OpenAI-compatible APIs, and GLM, MiniMax, and Kimi presets, each set by API key, base URL, and model name
- VM-level isolation: WSL2 on Windows or Lima on macOS runs every Bash command inside an isolated Linux virtual machine
- Path guard: file operations stay inside the chosen workspace folder even when no virtual machine is available
- Skills system: built-in PPTX, DOCX, PDF, and XLSX workflows, with custom skills created and deleted from the app
- MCP connectors: browsers, Notion, and other desktop services attach through the Model Context Protocol
- GUI operation: the agent takes control of desktop applications through computer use
- Remote control: commands and results relayed through Feishu (Lark) and Slack
- Multimodal input: files and images dragged and dropped into the chat input
- Trace panel: reasoning steps and tool calls displayed as they happen
- Local operation: the app runs on your machine and communicates only with the model API you configure
