BigQuery MCP Server: Set Up the Prebuilt Toolbox

The prebuilt BigQuery configuration in MCP Toolbox for Databases gives a local MCP client a documented route to BigQuery. This guide follows that route from credentials to a connection check. It does not compare tool counts with a separate managed service or assume that a hosted endpoint removes setup requirements.

Docs-derived; sources checked October 17, 2026. MCPFind did not authenticate to BigQuery or run a query for this article.

This guide matches the listing io-github-googleapis-mcp-toolbox, whose repository is googleapis/mcp-toolbox. Its setup uses a downloaded toolbox executable; do not substitute an unrelated npm or PyPI package.

Prepare your Google Cloud project

Use a project with billing enabled and the BigQuery API enabled. Ask the project administrator for the access needed for the intended datasets and operations; Google's guide identifies BigQuery User and Data Viewer roles or equivalent permissions. Configure Application Default Credentials for the environment running the client. Google's BigQuery Toolbox guide.

For local user credentials, Google's ADC setup documents the following command after installing the Google Cloud CLI:

gcloud auth application-default login

Follow the browser flow for the account with the intended access. ADC setup is separate from choosing a project in an application, so confirm which identity your local process will use. Local ADC instructions.

Install the Toolbox binary

Use the download section in the official setup guide to select a release for your operating system and CPU architecture. Save the binary to a location you will keep. On macOS or Linux, make it executable and check its version from that directory:

chmod +x toolbox
./toolbox --version

Configure Claude Code

Create or merge into .mcp.json in your project root:

{
  "mcpServers": {
    "bigquery": {
      "command": "/absolute/path/to/toolbox",
      "args": ["--prebuilt", "bigquery", "--stdio"],
      "env": {
        "BIGQUERY_PROJECT": "YOUR_PROJECT_ID"
      }
    }
  }
}

Replace both placeholders. On Windows, supply the actual executable path using JSON-escaped backslashes or forward slashes. This example uses the prebuilt configuration, so there is no tools.yaml to create. Google's client section documents these arguments and the project environment variable. Client configuration.

BIGQUERY_PROJECT selects the default execution project. Confirm which project will run and be billed for query jobs, and which datasets the ADC identity can read. Access to the execution/billing project does not grant access to every dataset; dataset IAM is a separate decision. A dataset in another project needs its own permissions. Review both with the administrator before execution. Google's BigQuery setup and permissions.

Verify metadata before querying data

Start Claude Code in the project, inspect the server in /mcp, and use a known, non-sensitive dataset for this docs-derived check:

1. List the available datasets, then the tables in that one dataset, and inspect one table's schema. Ask for metadata only, without reading rows or changing data.
2. Expect the known dataset and table identifiers, column names, and types. Compare them with the Cloud console using the same account before asking for SQL.
3. Review a proposed read-only query that names the fully qualified table, selects only necessary columns, and restricts the relevant partition or date range. Add a small row limit for the output, and review query cost controls: a LIMIT alone does not bound bytes scanned.
4. Confirm the execution/billing project, dataset permissions, and query before approving it. Expect only the reviewed columns and bounded result rows; compare a known sample with the source and inspect the job's usage before expanding.

Metadata success establishes a narrower check than query execution. Separate approval is required for writes, schema changes, or broader data access.

If the process fails to start, check the executable path and permissions. If it starts but access fails, check ADC, the selected project, and IAM permissions. Record the first failing operation rather than replacing the configuration with an unrelated server.

A successful connection does not establish that a generated query is correct, economical, or authorized. Browse database integrations for discovery; use the publisher's setup documentation for each server you actually install.