metabase-mcp-pro

# metabase-mcp-pro An agent-first **Model Context Protocol (MCP) server for Metabase** with a pluggable, table-scoped **query-guardrail layer**. It combines the lean ergonomics of Metabase's first-party MCP (guided MBQL builder, URI-addressed reads, pagination) with the full reach of the Metabase REST API (models, metrics, dashboards, admin — plus a generic API passthrough so nothing is out of reach), and adds the layer neither has: query guardrails that are **safe-by-default, explicitly overridable, and never silently filter**, plus an output sanitizer that strips connection secrets and trims token bloat. ## White-label No embedded brand, no environment defaults, no shipped table registry. Everything organization-specific (which tables are CDC, which need crawler exclusion, the vetted-query catalog) is supplied at runtime via env / config. ## Configure | Env var | Purpose | |---|---| | `METABASE_URL` | Metabase base URL (required; no default) | | `METABASE_API_KEY` | API key, sent as `x-api-key` | | `METABASE_MCP_GUARDRAILS_CONFIG` | path to a `guardrails.yaml` (see [`examples/`](./examples)) | | `METABASE_MCP_READONLY` | `1` to hard-disable all writes | ## Run ```bash uvx metabase-mcp-pro ``` Requires only `uv` — it installs the package from PyPI on demand. ## Use it in another project It's white-label, so any project reuses it by adding one MCP-server entry and supplying its **own** config. Drop this into a plugin manifest's `mcpServers` (or `~/.claude.json`): ```jsonc "metabase": { "command": "uvx", "args": ["metabase-mcp-pro"], "env": { "METABASE_URL": "https://metabase.your-org.com", "METABASE_API_KEY": "...", "METABASE_MCP_GUARDRAILS_CONFIG": "/path/to/your-project/guardrails.yaml" } } ``` Each consumer brings its own `METABASE_URL`, API key, and `guardrails.yaml` (start from [`examples/guardrails.example.yaml`](./examples/guardrails.example.yaml)) — nothing project-specific is baked into the package.