Skip to main content

​
Agent memory in any MCP client

turnlog mcp is a plain stdio MCP server. Any MCP-capable client can use it, not just the agent that wrote the sessions — and it serves every indexed agent’s history, so pointing Cursor at it gives Cursor access to your Claude Code and Codex work too.

​
The command

Every client needs the same thing: run npx turnlog mcp, talk to it over stdio.
Where that JSON goes depends on the client — Cursor takes it in its MCP settings, and most others use the same shape. If a client wants a single command line instead, it is npx turnlog mcp with no arguments. Installed globally? Use turnlog as the command and ["mcp"] as the args, which avoids the npx resolution step on every start.

​
Notes that bite

stdout is the protocol. In MCP mode every diagnostic goes to stderr — nothing is printed to stdout that is not JSON-RPC. If you wrap the command in a script, keep it that way or the client will fail to parse the handshake. It reads the same index as the UI. No separate configuration, no second copy. If the UI can find your sessions, so can the agent. It is read-only, permanently — see Claude Code setup for why that is a design decision rather than a default.

​
Checking it works

Run it in a terminal and it waits on stdin for JSON-RPC. That silence is success — it means the server started and is listening on the protocol channel. Ctrl-C to exit. For anything more, turnlog doctor reports what the index contains and whether it has drifted from what is on disk.