Setup guide

ChatGPT — MCP connector verified

Requires a paid ChatGPT plan (Plus or higher — Free has no custom-app support). Verified end-to-end on ChatGPT Plus: OAuth sign-in, board reads, live task updates, the one-shot plan import, and the session wrap.

1. In ChatGPT: profile → Settings → Plugins → enable Developer mode. (OpenAI renames this area often — it has also lived under Connectors / Apps; look for the developer toggle.)
2. Add a custom plugin with the server URL https://mcp.zimmboard.com/mcp and authentication OAuth.
3. ChatGPT opens the ZimmBoard sign-in page. Log in and click Allow access.
4. In a chat, enable ZimmBoard from the + / tools menu, then say “Load my zimmboard board and ground yourself.”

Alternative: a custom GPT with an Action

Prefer a dedicated GPT? In your board, open Connect AI → ChatGPT — it gives you the Action schema URL, a key, and ready-made GPT instructions (GPTs → + Create → Configure → Actions → Import from URL → Authentication: API Key, Bearer). Also requires a paid plan; Free gets neither path.

Le Chat (Mistral) — MCP connector verified

Works on Le Chat's free plan. Verified end-to-end: OAuth sign-in, board reads, live task updates, the one-shot plan import (nested sub-projects included), decision logging, and the session wrap.

1. At chat.mistral.ai, open the side panel (toggle at top-left), expand the Intelligence group, and click Connectors.
2. Click + Add Connector, switch to the Custom MCP Connector tab, name it ZimmBoard, and enter the server URL https://mcp.zimmboard.com/mcp.
3. Le Chat opens the ZimmBoard sign-in page — log in with your zimmboard account and click Allow access.
4. In a chat, open the Tools menu and tick ZimmBoard, then say “Load my zimmboard board and ground yourself.”

Cursor — MCP connector verified

Works on Cursor's free plan. Verified end-to-end, including the trickier paths: one-shot plan import with nested sub-projects, board files with surgical find-and-replace edits, and sync mode reconciliation.

1. Create Cursor's global MCP config. In Terminal:

   mkdir -p ~/.cursor && cat > ~/.cursor/mcp.json <<'EOF'
   { "mcpServers": { "zimmboard": { "url": "https://mcp.zimmboard.com/mcp" } } }
   EOF

   (Or per-project: the same JSON at .cursor/mcp.json in your project root.)
2. Fully quit Cursor (Cmd-Q) and reopen it.
3. Open Cursor Settings → Tools & MCP. Click zimmboard to start sign-in — your browser opens the ZimmBoard page; log in and click Allow access. You should see a green dot and 18 tools.
4. Open the chat with Cmd-L and switch the mode dropdown to Agent, then say “Load my zimmboard board and ground yourself.”

VS Code (GitHub Copilot) — MCP connector verified

Works with Copilot's free tier. Verified end-to-end: OAuth sign-in, board reads, the one-shot plan import with nested sub-projects, task updates, decision logging, and the session wrap.

1. Install the GitHub Copilot extension and sign in (the free tier includes agent mode).
2. Cmd-Shift-P → MCP: Add Server → HTTP → URL https://mcp.zimmboard.com/mcp → name it zimmboard → scope Global.
3. Open Copilot Chat and switch the mode dropdown to Agent.
4. Say “List my zimmboard boards” — the first tool use opens the ZimmBoard sign-in page. Log in, click Allow access, and you're connected.

Any AI — paste-in prompt or raw API

Any LLM that can make authenticated HTTP calls can use the board API with a Bearer key. In a board, open Connect AI → Any AI and it generates a single paste-in message with a fresh key built in — nothing to wire up. (Keys can be created read-only if the agent should never write.)

Under the hood it's two calls:

« start of session »
GET https://zimmboard.com/api/v1/boards/{id}/context
Authorization: Bearer YOUR_KEY

« end of session »
POST https://zimmboard.com/api/v1/boards/{id}/wrap
{"summary": "what I did", "session_tag": "Chat 12",
 "task_updates": [{"id": 3, "done": true}],
 "new_tasks": [{"project_id": 1, "text": "follow up"}],
 "carry_forward": ["next-session item"]}

The context response documents the full mid-session write API (add/check/edit tasks, log decisions, edit files, import plans), so your agent learns the rest by reading the board. So far this path is verified working with Claude; it should work with any agent that can send authenticated requests, and we'll note more agents here as they're tested.