transcripts-mcp lets one AI coding tool read another's history

transcripts-mcp reads the session transcripts that Cursor, Claude Code and Codex already write to disk and exposes them over MCP. Any of the three tools can then search the other two's history with fuzzy matching, plain-text search or regex, backed by a local full-text index. Semantic ranking is optional. Session browsing can be filtered by project or by date, and basic search works without building an index first.

That matters because switching tools has meant losing context. Start a debugging session in Claude Code, move to Cursor for the editor experience, and the reasoning from the first session is gone unless you copy it over by hand. This closes that gap by making the transcript itself queryable from wherever you're working.

It's precise about what it does and doesn't do. It returns transcript text for the current client to read, so it can recover prior reasoning and implementations. It does not restore a previous session's model state. You get the record, not a resumed conversation.

The design bet is that the transcript is the valuable artifact, not the running session. If that holds, the next step is a shared index across tools instead of three separate silos, which the README doesn't describe as a goal yet.