Skip to main content
Sidecar creates a managed BigQuery MCP connection from your existing BigQuery integration. It connects to Google’s server at https://bigquery.googleapis.com/mcp with the integration’s service-account identity. Sidecar manages short-lived Google access tokens and their OAuth renewal for Agent Console. You do not enter a separate MCP token or rotate it manually. Update the service-account key through the BigQuery integration when needed.

Prerequisites

  • Account Admin permissions in Sidecar
  • A connected BigQuery integration
  • The BigQuery API enabled in the projects agents should query
  • The integration’s service account granted mcp.tools.call, bigquery.jobs.create, and the required metadata and data-read permissions in each project
The current Sidecar custom-role instructions include mcp.tools.call. If your role was created earlier, add it in every project where agents should use MCP. Table reads also require bigquery.tables.getData and access to the selected datasets. See Google’s BigQuery MCP setup.

Network access

Sidecar’s discovery and readiness checks connect to Google’s endpoint from Sidecar. Agent Console MCP calls connect directly from Anthropic. If your Google Cloud policies restrict requests by source, ask your administrator to permit both access paths for the service account and projects. Use the outbound source addresses and your organization’s Google Cloud access controls. A successful query from Sidecar does not prove that an Anthropic-origin request is permitted.

Confirm the connection

  1. Save your BigQuery integration with the service-account JSON key.
  2. Open Integrations → MCP Servers and find the BigQuery row.
  3. Review the status and discovered tools. If setup is still pending, reload the page. Contact support if the row remains missing.
  4. If the row shows Auth error, review its status detail and check the service-account key and project permissions.
  5. Confirm the agent grant and job selection.
BigQuery MCP is managed from the BigQuery integration, even if the Add MCP dialog labels BigQuery as coming soon. Do not register the native endpoint as a custom server with a short-lived Google access token. Sidecar’s authenticated Check now and hourly readiness checks discover tools and execute SELECT 1 in each configured project. They confirm query-job execution from Sidecar, not read permission on every table or Anthropic’s access.

Verify Agent Console access

Start a new manual run after confirming the grant and job selection. Replace the placeholders with a project, dataset, and existing table agents may read:
In the run transcript, confirm successful BigQuery MCP calls for both queries. An empty table can return zero rows successfully. Complete any required approval. See runtime verification.

Available tools

Google’s server exposes read-oriented tools including list_dataset_ids, list_table_ids, get_dataset_info, get_table_info, and execute_sql_readonly. Queries run as the service account from your BigQuery integration. See the BigQuery MCP reference. The optional Cloud Logging settings in the warehouse integration guide support metadata ingestion. They do not create a Cloud Logging MCP server for Agent Console.

Alternative: self-hosted BigQuery MCP

If you need other tools, host a compatible MCP server behind an HTTPS endpoint that accepts a stable bearer token. Register it through the custom connector, including its network requirements for both Sidecar and Anthropic.

Troubleshooting

Check the status detail, the stored service-account key, and the required permissions in each configured project. Allow time for recent IAM grants to propagate, then select Check now. Update the key on the BigQuery integration’s Connection tab; manual MCP token replacement is unavailable.
Confirm the grant and job selection, then start a new run. Check Google Cloud access policies for Anthropic-origin requests and the service account’s permissions on the exact target table. Sidecar’s SELECT 1 readiness check does not test table access.
Grant the required permissions in every project agents query, not only the service account’s home project. Check the project named in the error and dataset-level access controls.