OpenWork: The Open-Source Claude Cowork Alternative (2026)

Every AI-agent ecosystem has the same bottleneck: your skills, MCP servers, and connected services live locked inside whichever agent you happened to open. Switch from Codex to Cursor and you start over. OpenWork wants to be the layer that travels with you. It’s an open-source desktop app — an alternative to Claude Cowork — but the part that got me is the MCP server, because that works in the agents you already have.

What OpenWork Actually Is

OpenWork (19,105 stars, +915 today — that’s real momentum, not hype) is a free desktop app for sharing AI workflows across macOS, Windows, and Linux. So the pitch in one line: add one OpenWork MCP to Codex, Claude Code, Cursor, or any compatible agent, and reuse the same skills, plugins, MCP connections, Google Workspace, and Microsoft 365 capabilities everywhere. Create something once, share it with teammates, or keep it for yourself.

But the part that sets it apart from a plain “workflow hub” is that the desktop app is optional. The README is explicit: “you can use OpenWork from the agent you already have.” That’s the whole game. It’s not another tool you have to learn — it’s a compatibility layer for the tools you already fight with.

Quick Start — I Actually Probed the OpenWork MCP

The fastest path is to paste a prompt into any command-capable agent and let it install OpenWork for you. But I wanted the wiring myself, so I hit the remote MCP endpoint directly:

curl -X POST https://api.openworklabs.com/mcp/agent \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize",
       "params":{"protocolVersion":"2025-03-26","capabilities":{},
                 "clientInfo":{"name":"test","version":"1.0"}}}'

Still, it answered HTTP 401 — which is exactly what you want from an MCP endpoint that’s meant to sit behind OAuth. The server’s live; it just wants you signed in. The endpoint exposes two tools once you’re authenticated: search_capabilities (finds what you can use) and execute_capability (runs it). Two functions. That’s the whole surface area.

Adding it to your client is a one-liner per agent:

# 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

Or drop a remote block into your opencode.json if that’s your driver. After adding, the client opens a browser so you can sign in and pick your organization.

Took me longer to read the docs than it would to wire it.

The Catch: “Open-Source” Doesn’t Mean Self-Hosted

Here’s the honest part. OpenWork is built around a hosted control plane — the MCP talks to api.openworklabs.com, and the Den admin interface manages teams, model providers, and published skills. “Open-source” and “self-hosted” are not the same thing here. If your whole draw is keeping your agent stack on your own infra, the remote MCP URL is a thing you’ll need to look at carefully. The repo’s TypeScript and the docs do cover local dev (a pnpm dev path that spins up Electron with a mock keychain), but the turnkey story leans on their cloud.

If you’re a solo dev or a small team that already trusts cloud SaaS for their agents, that’s fine. For the self-host-everything crowd, it’s a yellow flag worth reading the source for. If you’d rather keep that control plane on hardware you own, a DigitalOcean Droplet ($200 free credit for new users) is a cheap box to run your own MCP endpoints and agent services on (affiliate link).

OpenWork Verdict

OpenWork is worth a try if you’re juggling more than one agent and you’re tired of reconfiguring skills and MCPs every time you switch. Honestly, the two-tool MCP design is refreshingly minimal, and 915 stars in a single day says the pain it solves is real. My take: install it on your most-used agent, wire up one skill you actually care about, and see if portability beats the setup cost. The current version (v0.18.12, released July 30) is shipping fast and the team is clearly iterating. And if the hosted control plane ends up being the sticking point for you, standing up your own minimal agent stack on a low-cost Vultr instance ($100 trial credit) gives you the same reuse without handing the control plane to a third party.