Metabase MCP Server

Metabase ships an MCP server inside the product itself, served from /api/metabase-mcp on your own instance. That detail matters more than it sounds: queries run through Metabase's Agent API and semantic layer, and every tool call inherits the permissions of the person who authenticated. An analyst's agent sees exactly the collections, databases and rows the analyst sees — there is no separate service account quietly holding wider access.

The tool set is unusually complete for a BI integration. Read-only tools cover search (keyword or natural-language lookup of tables and metrics), construct_query and execute_query (build a query against a table or metric, then run it and get column metadata, row count and execution time), query with continuation-token paging at 200 rows a page and 2,000 total, and read_resource for reading collections, dashboards, databases, metrics, models, questions, schemas, tables and transforms by metabase:// URI.

Write tools go further than most: create_question, create_dashboard, create_collection, update_question and update_dashboard mean an agent can leave artefacts behind in Metabase rather than just answering in chat. execute_sql runs raw SQL and requires native-query permission on the target database; admins can switch it off instance-wide with the mcp-execute-sql-enabled setting.

The interactive tools are the fun part. visualize_query and render_drill_through render results as inline bar, line or table charts inside supporting clients — Claude Desktop and Claude on the web, Cursor and VS Code, and ChatGPT — which an admin enables per client (doing so adds that client's sandbox domains to Metabase's CORS allowlist). Command-line clients such as Claude Code need no allowlist entry at all.

An admin has to turn the server on under Admin → AI → MCP, and AI features must be enabled on the instance, though you do not need to configure an AI provider. Metabase is open source, so this works on a self-hosted instance as well as Metabase Cloud.