Connecting an AI client over MCP

MCP (Model Context Protocol) lets an external AI client — such as Claude Desktop, Cursor, or any MCP‑compatible app — securely use your data and tools from this platform. You generate a small config here, paste it into your client, and the client can then query your connected sources through the assistant.


Step‑by‑step: connect your client

  1. Generate the config. In the chat settings, click the 🔌 Generate MCP client config button. A window opens with a ready‑to‑use JSON snippet.
  2. Copy the config. Use the copy button in the top‑right of the snippet.
  3. Open your client's config file and paste it in: - Claude Desktop / Claude Code: ~/.claude/settings.json - Cursor: ~/.cursor/mcp.json - Other MCP clients: wherever the app stores MCP server entries. - If the file already has an "mcpServers" section, add this entry alongside the existing ones (don't replace them).
  4. Save the file and restart the client. Most clients pick up MCP servers only at startup.
  5. Check the tools. In your client you should now see the tools provided by this platform. Ask the assistant something that uses your data — it will call the tools automatically.

Quick way: just ask the assistant

You don't have to edit a config file by hand. You can paste the copied JSON straight into the chat and tell the assistant something like:

"Connect to my MCP server using this config" — then paste the JSON.

The assistant will read the endpoint and token from the config and connect for you.


For the AI: how to connect correctly

This server speaks the Streamable HTTP transport of MCP. If you're an AI/agent setting up the connection, follow these rules — getting session_id wrong is the most common mistake:

  1. Initialize first. Send the MCP initialize request to the endpoint (the url in the config) with the Authorization: Bearer <token> header.
  2. Read the session id from the response. The server returns it in the Mcp-Session-Id HTTP response header during initialization. Do not invent it and do not read it from the URL.
  3. Echo it on every later request. Send that exact value back in the Mcp-Session-Id request header on all subsequent calls (tools/list, tools/call, notifications, etc.). Requests without the correct session header are rejected.
  4. Keep the Bearer token on every request too — both headers are required together.
  5. One session per connection. Reuse the same Mcp-Session-Id for the whole conversation; only re-initialize (and get a new id) if the session expires or the server returns a session error.

In short: initialize → grab Mcp-Session-Id from the response header → resend it (plus the Bearer token) on every following request.


Security & token


FAQ / troubleshooting

My client doesn't show any tools. Make sure you saved the config and fully restarted the client. Tools are loaded only when the client starts. Also confirm you pasted the entry inside "mcpServers" and the JSON is still valid (no missing commas or braces).

I get a 401 / "unauthorized" or "access denied" error. The token has likely expired because you logged out. Generate a new config here and replace the old entry, then restart the client.

The connection drops or tools disappear after a while. An idle MCP session is closed automatically after about 30 minutes. Just send a new request from your client — it will reconnect. If it doesn't, re‑generate the config.

Can I use the same config on two machines? Yes, but both use the same token. If you regenerate the config (e.g. after re‑login), update it on every machine.

Is my data sent anywhere else? No. The AI client talks only to this platform using your token, and tool calls run against the sources you have connected.