A BigQuery MCP server lets an AI explore datasets, inspect schemas and run analytical SQL against your warehouse. Google supports two routes in 2026 — a managed remote server and the open-source MCP Toolbox for Databases. Here is how to set each up.
Option A — MCP Toolbox for Databases (open source)
Google's open-source mcp-toolbox (formerly Gen AI Toolbox) is a standalone MCP server that connects to BigQuery and many other databases. Authentication uses Application Default Credentials, so first run:
gcloud auth application-default login
Then run the toolbox pointed at a BigQuery source. The server listens locally (default port 5000) and you connect your client to it over HTTP. The toolbox repo documents the BigQuery source config (project + dataset). This route gives you the most control and works fully self-hosted.
Option B — managed remote BigQuery server
Google also offers a fully managed, remote BigQuery MCP server (in preview as of 2026). You add it as a remote MCP connector in clients that support OAuth/HTTP servers and authorise with your Google account — nothing to host. Best when you want zero infrastructure.
Verify
Ask: "List the tables in dataset analytics and describe events." Real schema back means you are connected.
Safety: read-only and least privilege
Grant the credentials only the BigQuery roles needed to read the datasets in scope — typically BigQuery Data Viewer plus BigQuery Job User to run queries. Do not hand an AI write or admin roles on your warehouse. This is the least-privilege principle from MCP security best practices.
Common problems
Reauthentication is required/ 401 — ADC expired; rerungcloud auth application-default login.Permission deniedon a dataset — the credential's role does not cover that dataset.- Port 5000 in use (Option A) — start the toolbox on another port and update the client URL.
Going further
For Snowflake, the same toolbox applies — see Snowflake MCP setup. Comparing warehouses? Read MCP for BigQuery vs Snowflake and best MCP server for a database.