Metabase MCP server

What it does

This is not a separate package: the MCP server ships inside Metabase itself, the open-source BI application, and is served by your own instance at https://<your Metabase host>/api/metabase-mcp once an admin switches it on. An AI client connects to that address, signs in through Metabase's own OAuth server, and then acts as the person who signed in, limited to what that person can already see and do in Metabase.

It has 23 tools. Read tools search and browse collections, tables and fields, fetch questions, dashboards and other items, look up glossary terms and filter values, and run saved questions, structured queries built from table and field ids, or raw SQL. Results come back as text or, in clients that can render them (Claude Desktop and web, Cursor, VS Code, ChatGPT), as an inline chart with a link to explore the question in Metabase. Write tools create and edit questions, dashboards, documents, collections, metrics, measures and segments, bookmark or duplicate items, and set up alerts and dashboard subscriptions that send data by email or to Slack.

The client brings the model; Metabase needs no AI provider configured for this, and MCP calls do not count against a Metabot AI connection. Each tool belongs to one OAuth scope (content read, content write, query run, SQL run, delivery write), shown on the consent page. A fact the docs do not spell out: the source still serves the older /api/mcp path as an alias for back-compatibility (src/metabase/mcp/paths.clj), which is the address the MCP registry entry lists; current docs give /api/metabase-mcp.

What it needs

A running Metabase instance (Metabase Cloud or self-hosted) and a Metabase account on it. An admin must turn on the MCP server under Admin > AI > MCP, and AI features must not be disabled under Admin > AI. The client must support remote streamable HTTP servers with OAuth and dynamic client registration. Inline charts in browser-based clients need that client's toggle (or its origin) enabled by an admin.

Limits

Only works against a Metabase instance; there is no local stdio package. Structured queries, saved questions and SQL return 100 rows per call by default, up to 2000 with row_limit (paging with a cursor for more); search returns 20 results by default, up to 50; parameter values 100 by default, up to 1000. duplicate_content copies questions, dashboards and documents only, not models or metrics. Alerts can be paused but not deleted through the tools. Clients that cannot render inline charts, such as Claude Code, do not see the visualize_query tool. Clients connected before Metabase 0.64 must sign in again after upgrading. Usage analytics for MCP calls are on Pro and Enterprise plans only.

Who it suits

Suits teams that already run Metabase and want an assistant to answer questions from their existing models, metrics and permissions, or to draft questions and dashboards, without handing out an API key: sign-in is per person through OAuth and every call is bounded by that person's Metabase permissions. It is published by Metabase and needs no extra process. Pick one of the community Metabase servers instead only if your instance cannot be upgraded to a version with the built-in server or your admin will not enable it; check each one's own page for how it signs in and what it can do.