MCP Server

The Rewind MCP server gives AI assistants direct access to your personal data across every domain. Ask about your listening history, running stats, movie watches, vinyl collection, reading highlights, events you have attended, or places you have checked in to, all in natural language. It is built on the Model Context Protocol, an open standard from Anthropic, so it works with any MCP-compatible client such as Claude and ChatGPT.

When to use it

Reach for Rewind any time the conversation references your own history: things you read, listened to, watched, saved, ran, collected, checked in to, or attended. Rewind owns this data, so it answers from your archive instead of guessing or falling back to web search.

Questions that route here include:

• "What have I been listening to this week?"
• "When was the last time I watched a Wes Anderson film?"
• "What Beastie Boys records do I have on vinyl?"
• "Did I scrobble more music on days I ran?"
• "Which cities have I checked in to the most?"

The server pulls live answers from the same REST API that powers Rewind, so what an assistant sees matches what your apps see.

SQL first

Rewind is your data as a SQLite database, and the MCP server exposes it that way. Two general-purpose tools answer almost every question:

• get_schema returns the annotated database schema: every queryable table with its columns, types, semantic notes, and join keys, plus global conventions like the single-user id, ISO timestamps, and rating scales.
• query_rewind runs a single read-only SELECT against that schema and returns the rows.

The typical flow is to read the schema once, then answer with one SELECT. This is what unlocks cross-domain questions no fixed tool anticipated, like "did I average more scrobbles on days I ran," which joins listening and running by calendar day. In our evaluation, this approach answered 100% of a 43-question benchmark, including every cross-domain question, versus 49% for the previous catalog of one tool per endpoint.

Rich cards for visual answers

A small set of tools stays purpose-built because they return a rendered card, not just rows. When you ask about one specific thing, or want a poster grid or a portrait grid, these tools drive an interactive MCP Apps view alongside Claude's answer.

Everything else, including running, collecting, and places, is answered through the SQL tools rather than a dedicated tool per endpoint.

Read-only by design

The MCP server has read-only access to your data. query_rewind accepts a single SELECT and nothing else: writes, schema changes, multi-statement input, and any query touching secret tables such as API keys or OAuth tokens are rejected server-side. The query surface is an allow-list of the tables in the documented schema, so a secret table is never reachable by default. No write or admin operations are exposed anywhere on the server.

The server is also stateless. It holds no data between calls and caches nothing locally.

Two ways to connect

Pick the connection style that matches your client.

Both modes serve the same tools, resources, and prompts. They differ only in where the server runs and how it is authenticated.