OpenWork

OpenWork is a free, open-source desktop app for sharing AI workflows across agents and teammates. It works as an open-source alternative to Claude Cowork and Codex, and it targets developers and teams who already use multiple AI coding agents (Codex, Claude Code, Cursor, OpenCode) and want to reuse the same skills, MCPs, and connected services between them instead of rebuilding setups per tool.

Instead of locking workflows into one client, OpenWork exposes a remote MCP server. Any compatible agent can connect to it, search for available capabilities, and execute them. A desktop app exists for teams that want a dedicated workspace, but it's optional: you can keep using the agent you already have and just add the OpenWork MCP endpoint.

Key features

• Cross-agent MCP: one MCP server (https://api.openworklabs.com/mcp/agent) works with Codex, Claude Code, Cursor, OpenCode, and any MCP-compatible client.
• Two-tool interface: search_capabilities finds what's available to a user, execute_capability runs it, keeping the surface area small.
• Shared skills and connections: create a skill, MCP, or service connection once and reuse it across machines, tools, and teammates.
• Desktop app: a dedicated macOS, Windows, and Linux workspace for running and managing workflows, with install available via download or an agent-run install prompt.
• OpenWork Den admin console: a control plane for organizations to provision inference, manage teams and access, set desktop policies, and publish skills/plugins through marketplaces.
• Anthropic-compatible plugin import: pull in existing Anthropic-compatible plugins so their skills and remote MCPs become available through the OpenWork MCP.
• Google Workspace and Microsoft 365 capabilities: connected services surface as capabilities agents can search and execute.
• Headless web mode: run the OpenWork UI in a browser against a local openwork-server, with isolated config, stable dev bearer tokens, and same-origin proxying for Den Cloud calls.

Ideal use cases

OpenWork fits teams or individuals juggling several AI coding agents who are tired of reconfiguring the same MCPs and skills in each tool. If you built a useful workflow in Claude Code and want it available in Cursor or Codex without copy-pasting config, this solves that directly. It also fits organizations that need to control which model providers, skills, and plugins different teams can access, since Den provides the admin layer for that.

It's a good fit for teams standardizing on shared connected services (Google Workspace, Microsoft 365) across multiple AI tools, or for anyone who wants a single place to publish and version internal automation skills.

It is not a good fit if you use a single AI agent exclusively and have no need to share workflows across tools or people, since the cross-agent MCP layer adds complexity you won't use. It's also not the right choice if you need a fully offline, air-gapped setup with no remote MCP server dependency, given that the default flow relies on a hosted api.openworklabs.com endpoint and Den for organization sign-in.

Installation

The fastest path is to let an existing AI agent install it. Paste this prompt into Claude Code, Cursor, Codex, ChatGPT, or any agent that can run commands on your machine:

Install OpenWork on my computer, set up my first workspace, and open it ready to use. Follow the steps in https://openworklabs.com/start.md?v=hero

Alternatively, download the desktop app directly from openworklabs.com/download.

To use OpenWork from an existing agent instead of the desktop app, add the MCP server:

# Codex
codex mcp add openwork --url https://api.openworklabs.com/mcp/agent

# Claude Code
claude mcp add --transport http openwork https://api.openworklabs.com/mcp/agent

For OpenCode, add this to opencode.json:

{
  "mcp": {
    "openwork": {
      "type": "remote",
      "enabled": true,
      "url": "https://api.openworklabs.com/mcp/agent",
      "oauth": {}
    }
  }
}

Any other MCP-compatible client can point directly at https://api.openworklabs.com/mcp/agent. After adding the MCP, the client opens a browser for sign-in and organization selection.

For local development on the repo itself, use pnpm dev for a single checkout, or pnpm dev:worktree to run multiple git worktrees simultaneously with isolated dev profiles. A headless, browser-only mode without Electron is available via pnpm dev:headless-web, useful for testing the UI against a local openwork-server without the desktop shell.