// the find
ClickHouse/mcp-clickhouse
Connect ClickHouse to your AI assistants.
An MCP server that lets AI assistants query ClickHouse databases via natural language. Built by the ClickHouse team on top of FastMCP, it exposes a small set of tools for schema discovery and SQL execution. Most useful if you're already running ClickHouse and want Claude or Cursor to interrogate it directly.
The two-tier write protection is well-designed — reads are safe by default, writes require one flag, destructive ops require a second explicit flag, which is exactly the right model for an AI client that might hallucinate a DROP. The chDB integration is a genuine differentiator: you can query Parquet files, URLs, or local data through an embedded ClickHouse engine without standing up a server at all. Auth for HTTP/SSE transports is required by default rather than opt-in, which is the right call. The middleware hook system for per-request connection overrides is a clean escape hatch for multi-tenant routing without forking the server.
The tool surface is minimal — there's no way to explain a query plan, inspect running queries, or see cluster topology, so an AI assistant will hit walls fast on anything beyond basic querying. The `list_tables` pagination is token-based but there's no equivalent for column metadata at scale; schemas with thousands of columns per table will overflow context. The chDB integration lives behind an optional extra and a separate tool name, which means an AI doesn't naturally know when to prefer embedded vs remote — the client has to configure this correctly up front. There are no streaming results: a long-running analytical query blocks the MCP connection for its full duration, and the default 30-second query timeout will surprise anyone used to ClickHouse handling multi-minute scans.