Coworker

What it does

Coworker is an open-source AI agent platform you can self-host or deploy to Railway. It provides an Electron desktop app and server backend that lets you interact with an AI team member capable of handling tasks, answering questions, building internal tools, and managing workflows — all via a chat interface. It ships with a full MCP integration layer, meaning you can visually manage MCP servers, discover new ones from a built-in registry, and let Coworker use them as tools.

Key features

• MCP UI — visual dashboard to add, configure, and monitor connected MCP servers
• App builder — Lovable-like UI for generating and maintaining internal dashboards; agents commit changes via git
• Skills marketplace — browse and install community skills from ClawHub and skills.sh directly from the chat
• A2A protocol — exposes an Agent-to-Agent endpoint so other AI agents can discover and delegate tasks to Coworker
• Scheduled tasks — cron-powered recurring AI workflows via Inngest
• Multi-provider AI — switch between OpenAI, Anthropic, Google Gemini, NVIDIA, Groq, Kimi, or any OpenAI-compatible endpoint from settings

Installation

Docker (recommended):

docker compose up

Self-host (Bun):

cp .env.example .env   # add API keys
bun install
bun run dev            # backend at http://localhost:4111
# In another terminal:
cd app && bun install && bun run dev

Deploy to Railway: One-click deploy available at railway.com/deploy/coworker. Includes Inngest, Postgres, Redis, and optional Tailscale networking.

Claude Desktop:

{
  "mcpServers": {
    "coworker": {
      "command": "bunx",
      "args": ["coworker"],
      "env": { "SERVER_URL": "http://localhost:4111" }
    }
  }
}

Supported hosts

Coworker is itself an MCP client that connects to other MCP servers. It also exposes an MCP server endpoint compatible with Claude Desktop and any MCP-capable client.