MotherDuck MCP

MotherDuck MCP gives an agent a SQL interface to DuckDB, whether the data sits in the MotherDuck cloud warehouse, a local DuckDB file, an in-memory session, or files in object storage. The server is deliberately thin: a single query tool that accepts DuckDB SQL and returns results, which means everything DuckDB itself can do is reachable, including reading Parquet, CSV and JSON straight from a path or an S3 bucket with no load step. That design suits exploratory analytics. An assistant can list databases and schemas to orient itself, describe a table, then write and refine the aggregate an analyst actually asked for, iterating on the SQL when a column name or a join turns out wrong. Querying files in place also makes it useful for one-off work on data that was never loaded into a warehouse: point the agent at a folder of exports and ask it to profile them. The server runs read-only by default and requires an explicit flag to enable writes, which is the right default for agent access to a warehouse. Connections can be switched without restarting, so one client can move between a scratch local file and a shared MotherDuck database. It fits data teams who want conversational analysis on top of DuckDB, and engineers who want an agent to draft and validate SQL against the real schema rather than an imagined one.