What MCP gives you
The integration layer in Comuna is built on MCP (Model Context Protocol), an open standard. That decision pays off three ways:
- One server, many clients. Comuna doesn't ship a Claude integration and a ChatGPT integration as separate code. There's one MCP server at
api.comuna.work/mcp. Any client that supports MCP can use it. - Open, not proprietary. MCP is published by Anthropic, supported by OpenAI, and being adopted across the ecosystem. New clients show up; the server stays the same.
- No vendor lock-in for you. Today you use Claude. Tomorrow you switch to ChatGPT (or use both at once). Same workspace, same permissions, same data — only the chat surface changes.
Available now
Claude (Anthropic)
First-class. Connect in Settings → Connectors → Add custom connector. Free Claude plan works. Full MCP tool support.
ChatGPT (OpenAI)
First-class. Requires Plus / Pro plan (OpenAI gates the MCP connector UI to paid tiers). Settings → Apps → Developer Mode → Create.
The MCP server itself
If you want to wire Comuna to a custom client, an agent framework, or a script: the server speaks standard MCP over HTTP with OAuth2. ~80 tools, REST mirror at /api/v1/* if you need plain JSON. Docs and tool catalogue inside.
Coming next
- Gemini (Google) — the underlying MCP server already supports it. Waiting for Google's MCP connector UI to GA in consumer Gemini.
- Cursor — same story. The MCP server supports it; we want Cursor's UI to stabilise before promoting it.
If you build a custom client today, the MCP server is fully documented — you don't have to wait for us.
Permissions, cost, privacy (the short version)
- Permissions: the AI inherits whatever permissions you have, scoped per board. You can further restrict it from the AI Coworker permissions panel.
- Cost: zero on Comuna's side, forever. You pay your AI vendor (Claude / ChatGPT) directly via your own subscription. We don't proxy tokens; we don't take a cut.
- Privacy: when the AI reads a card on your behalf, that card is sent to your AI vendor via MCP for that specific call. Comuna doesn't keep a copy, doesn't replay, doesn't train.
Per-client details on the pages above.
Honest about the limits
MCP is pull, not push. Comuna can't make any AI client act on its own — the AI only runs when you (or a scheduled prompt in its client) trigger it. We are explicit about this on the AI coworker page because pretending otherwise breaks the trust contract.
Pick a client, connect it, leave your AI work at the desk, give it a daily scheduled prompt to check Comuna every morning.