Summary

OpenCowork is an open-source, local-first desktop AI agent runtime designed for evaluating and prototyping browser automation, business workflows, and AI-driven tasks. It empowers builders, researchers, and product teams to transform business goals into executable, reusable, and auditable workflows, offering a privacy-conscious alternative for local execution. The project prioritizes observable, recoverable, and repeatable task execution, turning successful runs into reusable templates for complex operational scenarios.

How It Works

The core approach centers on a local desktop agent runtime that orchestrates multi-step tasks. Users describe a goal, and the agent operates local tools and websites under human oversight. Successful task runs are persisted as inspectable records and can be saved as parameterized, reusable workflow templates. The architecture is evolving towards a distinct Agent Runtime API, supporting execution adapters for browsers, desktops, and MCP (Machine Communication Protocol) clients/servers, with a shared protocol layer for core runtime services like lifecycle management, approval, and tracing. This design facilitates hybrid computer-user automation (CUA) by combining DOM-first browser automation with visual execution fallbacks and approval flows.

Quick Start & Requirements

• Installation: Clone the repository, run npm install, and npx playwright install chromium.
• Prerequisites: Node.js 20 or 22 LTS, npm 9+, Python 3.8+, and a valid LLM API configuration in config/llm.json. Specific OS prerequisites include bash, brew (macOS), or apt (Ubuntu/Debian).
• Setup: Requires configuring LLM API keys and potentially local tool integrations.
• Documentation: USER_GUIDE.md, docs/ARCHITECTURE.md, docs/ROADMAP.md.

Highlighted Details

• Hybrid CUA: Integrates DOM-first browser automation with visual execution fallback and human approval flows.
• Task Persistence: Task runs are saved with state, results, artifacts, and reusable context.
• Workflow Templating & Scheduling: Successful tasks can be saved as templates and executed on a schedule.
• MCP Interoperability: Supports both client (connecting to external tools) and server (exposing capabilities) roles for MCP.
• Feishu Integration: Enables task submission, progress updates, and file delivery via Feishu messages.

Maintenance & Community

The project is actively evolving, with recent releases focusing on scheduler dependability and execution handling. Community engagement is facilitated through GitHub Issues and Discussions.

• Releases: https://github.com/LeonGaoHaining/opencowork/releases
• Issues: https://github.com/LeonGaoHaining/opencowork/issues
• Discussions: https://github.com/LeonGaoHaining/opencowork/discussions
• Website: https://opencowork.me

Licensing & Compatibility

Licensed under the Apache-2.0 license. This permissive license generally allows for commercial use and integration into closed-source projects, with standard attribution requirements.

Limitations & Caveats

OpenCowork is explicitly positioned as an open-source evaluation and development project, not a production-certified RPA suite, hosted SaaS platform, or generic chatbot. It is best suited for local experimentation and prototyping. Users must treat the agent as a trusted automation tool with operating privileges, recommending its use on dedicated, isolated systems to mitigate risks of accidental operations, credential exposure, and data leakage. Sensitive configurations and API keys must be kept local and not committed.