OpenCowork is a local-first desktop AI Agent Runtime for teams that need real execution, not another chatbot. It operates browsers and local tools, connects MCP and skills, records auditable task runs, and turns successful sessions into reusable templates.
OpenCowork focuses on repeatable business outcomes: local execution, observable runs, and workflow reuse across browser-heavy and tool-connected tasks.
Use your local browser, files, tools, model configuration, and private deployment boundary instead of sending every workflow through a hosted SaaS runtime.
Execute browser actions, CLI steps, MCP tools, skills, IM file workflows, and scheduled jobs through one approval-aware task lifecycle.
Review the result, artifacts, screenshots, traces, and diffs, then save successful runs or whole sessions as templates that can be rerun and scheduled.
OpenCowork is designed for builders, researchers, product teams, and early enterprise pilots that want local execution, browser-first automation, MCP interoperability, workflow templates, and auditable task infrastructure in one open-source system.
Run multi-step work on a desktop agent that can operate websites, call local tools, use installed skills, and produce files or structured outputs.
Persist task state, results, screenshots, artifacts, run links, runtime traces, and reusable context so useful work remains inspectable after the live run ends.
Save proven runs or completed session workflows as templates, trigger them from chat, scheduler, IM, or MCP-oriented flows, and repeat work without rebuilding the process.
OpenCowork is strongest where a team repeatedly opens websites, processes files, connects internal tools, and needs a result that can be audited or reused.
Collect public information, compare competitors, monitor pricing, and generate lead lists or research briefs.
Output: brief, spreadsheet, PPT outlineReceive files through IM, run OCR or image analysis, apply repeatable rules, and return generated result files.
Output: cleaned file, OCR result, reportOperate non-standard dashboards, forms, approval consoles, and long-tail browser workflows with human oversight.
Output: completed operation, audit traceConnect local scripts, internal services, MCP servers, and skills into reusable private workflow packages.
Output: template, run record, artifactsOpenCowork combines browser execution, Hybrid CUA, desktop workflows, MCP tools, reusable templates, session management, IM intake, skills, and human oversight so one system can finish work and make it repeatable.
Navigate, click, type, wait, search, and extract in a headed browser for workflows that need real interaction instead of simulated demos.
Use DOM-first browser automation with visual execution fallback, approval-aware recovery, takeover paths, and persisted visual traces.
Preserve final summaries, structured data, generated files, screenshots, artifact links, runtime trace status, run details, and follow-up actions in a reviewable result surface.
Convert successful runs or active sessions into templates, schedule recurring work, and receive or deliver file-driven tasks through Feishu messaging workflows.
Install, update, uninstall, and run reusable capability modules, call local commands, handle files, and generate assets that remain useful outside the chat session.
Connect local stdio or remote streamable-http MCP servers, and expose selected OpenCowork capabilities through a standard MCP endpoint.
The product now includes the first shared local Agent Runtime baseline used by Electron, Scheduler, IM, MCP, and future clients through one task lifecycle.
Install Node.js, npm, Python, Playwright Chromium, configure your model locally, and launch the desktop app.
# Install requirements
# macOS
brew install node python
# Ubuntu / Debian
sudo apt update
sudo apt install -y nodejs npm python3 python3-pip
# Windows PowerShell
winget install OpenJS.NodeJS.LTS
winget install Python.Python.3.12
# Verify requirements
node -v
npm -v
python3 --version
# Clone the repository
git clone https://github.com/LeonGaoHaining/opencowork.git
cd opencowork
# Install dependencies
npm install
npx playwright install chromium
# Configure your model
# Create config/llm.json
# Keep config/ local and never commit credentials
# Start the desktop app
npm run electron:devOpenCowork is a local-first desktop AI agent runtime. It can operate a headed browser, call local tools, connect to MCP servers, process files, run scheduled workflows, and integrate with IM systems such as Feishu. Because it can perform real actions in a local desktop environment, use it as a trusted automation tool with operating privileges, not as a sandboxed chatbot.
To reduce accidental operations, credential exposure, and data leakage, run OpenCowork on a dedicated AI automation device, virtual machine, or isolated system account whenever possible.
Use trusted machines, networks, and operating system accounts.
Review agent actions before allowing access to sensitive websites, files, credentials, or production environments.
Keep `config/llm.json`, `config/feishu.json`, API keys, tokens, cookies, databases, and private artifacts local.
Verify scheduled tasks and reusable templates before running them unattended.
Report sensitive findings through GitHub Security Advisories instead of public issues.
OpenCowork is an Apache 2.0 open-source desktop AI work system with public releases, refreshed docs, issue templates, contributor guidance, and an active local Agent Runtime for real automation instead of demo-only agent experiences.
