Google offers two official MCP paths to BigQuery, and there’s a third option that doesn’t involve MCP at all. The right choice depends on your client application, your team setup, and what you actually need the AI to do with your data.
The Three Options at a Glance
| Remote MCP Server | MCP Toolbox | bq CLI (no MCP) | |
|---|---|---|---|
| Infrastructure | Managed by Google | Self-hosted binary | Already installed with gcloud |
| Authentication | OAuth (Claude Desktop) / manual tokens (Claude Code) | Application Default Credentials | Regular gcloud auth |
| Tools available | 5 basic tools | 5 basic + forecasting, catalog, custom queries | Full bq command set |
| Custom queries | No | Yes, via tools.yaml | Arbitrary SQL |
| Multi-database | BigQuery only | BigQuery, Cloud SQL, Spanner, others | BigQuery only |
| Setup effort | Minimal | Moderate | Zero |
| Best for | Claude Desktop quick exploration | Claude Code, team workflows, production | Interactive ad-hoc work in Claude Code |
Decision by MCP Client
Claude Desktop users: Start with the Remote Server. OAuth is handled automatically, setup is trivial, and the five built-in tools cover most exploration needs. Upgrade to the Toolbox when you need custom parameterized queries or multi-database support.
Claude Code users: Skip the Remote Server. Its token expiration (roughly hourly) makes it impractical for regular use — you’d spend more time refreshing tokens than querying data. Use the Toolbox, which handles authentication through ADC without manual intervention.
When You Don’t Need MCP at All
Before committing to MCP setup, consider whether the bq CLI is simpler. Claude Code can run bq query, bq ls, and bq show commands directly. No server to configure, no restart cycle when credentials change, no MCP overhead.
The CLI uses your regular gcloud credentials, requires no server restarts after auth changes, and gives clearer error messages. For interactive exploration — “What’s in this table? Run this query. Now try this one.” — the CLI is often the fastest path.
MCP adds value in specific scenarios where the CLI falls short:
Custom parameterized queries. The Toolbox’s tools.yaml lets you define exactly which queries the AI can run with which parameters. This is much safer than arbitrary SQL and useful when building AI workflows that should only access specific data patterns.
Multi-database workflows. If you work across BigQuery, Cloud SQL, Spanner, or other databases, the Toolbox provides a unified interface rather than switching between CLIs.
Structured tool outputs. MCP returns structured data that Claude can parse more reliably than CLI output. For automated pipelines where you need predictable response formats, MCP reduces parsing errors.
Audit requirements. Every MCP tool call creates a structured log entry. If compliance requires audit trails of AI data access, MCP provides this natively.
Decision by Use Case
Quick data quality checks (“Count null values in the email column,” “Find duplicate customer_id values”) — The bq CLI works perfectly. The queries are simple, the results are small, and the overhead of MCP isn’t justified.
Ad-hoc reporting (“Monthly sales summary for 2025 by product category”) — Either the CLI or MCP works. MCP’s structured output is slightly nicer for complex results, but the CLI is faster to get started with.
Constrained AI workflows (“The AI should only query these specific tables with these specific parameters”) — The Toolbox with custom parameterized queries is the only option that provides this level of control.
Team standardization (“Every developer should have the same AI-accessible queries”) — The Toolbox with a shared, committed tools.yaml and project-scoped .mcp.json. The CLI has no equivalent for standardizing query patterns across a team.
Cross-database analysis — The Toolbox, configured with multiple sources. The CLI is BigQuery-only.
Decision by Team Size
Solo practitioner: Start with the bq CLI. Add the Toolbox when you want parameterized queries or find yourself repeatedly writing the same setup instructions for AI queries.
Small team (2-5): The Toolbox with a project-scoped .mcp.json and shared tools.yaml. The setup overhead pays off quickly when multiple people benefit from the same curated query interface.
Larger team with compliance needs: The Toolbox with service account authentication, custom parameterized queries to restrict access patterns, and audit logging. The cost control measures become especially important at scale.
The Complementary Approach
These options are not mutually exclusive. A practical setup might use:
- The bq CLI for quick, interactive exploration during debugging
- The Toolbox with custom tools for structured, repeatable queries
- The Remote Server in Claude Desktop for casual browsing
The choice is not permanent — start with whatever reduces friction most and add complexity when the need arises.