MCP Toolbox for Databases (BigQuery and more)

Install

BigQuery (stdio, pinned)

npx -y @toolbox-sdk/server@1.13.0 --prebuilt bigquery --stdio

Set BIGQUERY_PROJECT and have Google Application Default Credentials available. Set BIGQUERY_READONLY=true unless the agent needs to write.

Other databases

npx -y @toolbox-sdk/server@1.13.0 --prebuilt=postgres --stdio

Swap postgres for snowflake, clickhouse, mysql, spanner, etc. Each needs its own connection variables — see the prebuilt tools reference.

A single Go binary (also on npm as @toolbox-sdk/server) that turns a database connection into MCP tools. With --prebuilt bigquery you get execute_sql, dataset and table listing and metadata, catalog search, ask_data_insights, analyze_contribution (what drove a metric change) and forecast, which runs BigQuery’s AI.FORECAST over a time series. The same flag works for Snowflake, ClickHouse, Postgres, MySQL, Spanner, Oracle, MongoDB and others.

When to use it

Ask questions of a warehouse in plain English and see the SQL that answered them.
A quick time-series forecast in BigQuery without setting up a forecasting library.
One MCP server across several databases, or custom tools with fixed, parameterised SQL.

Before you install

Lock it down first. Set BIGQUERY_READONLY=true (or writeMode blocked/protected), restrict allowedDatasets in a custom config, and set BIGQUERY_MAXIMUM_BYTES_BILLED so a bad query can’t run up a large scan bill. Use a service account or role with only the datasets the agent needs. Google also offers managed remote MCP servers for its databases if you’d rather not run this yourself.

Limits

Forecasts are statistical projections with uncertainty; read the prediction intervals, not just the point estimate. Agent-written SQL can be wrong while still running — check joins and filters on anything you report.