BigQuery MCP server optimized for quick navigation of larger projects and datasets.
BigQuery MCP has 9 well-named tools with clear descriptions and documented input schemas. However, output schemas are not documented in the source code, parameter descriptions lack detail on constraints and valid ranges, and no error handling guidance is visible. Tool names follow verb_noun convention (list_*, get_*, search_*, run_query) and are action-oriented. Descriptions are moderate length (60-150 chars) but lack specificity about when to use each tool vs alternatives. All 9 tools have input parameters with type declarations and basic descriptions, but parameter descriptions do not state constraints, ranges, or expected formats. Security model appears sound (read-only operations, no secrets in parameters), and the tool chain is well-designed (e.g., list_datasets → get_dataset → list_tables → get_table → get_table_schema, or search_tables → vector_search). No visible error handling patterns, output schema documentation, or parameter constraint validation in the provided code.
Generate embeddings using a BigQuery ML model and perform semantic search in one operation
Get detailed information about a specific dataset including its tables, description, and metadata
Get detailed information about a table including schema, sample data, and column statistics
Get the schema of a table with detailed field information
List all datasets in the BigQuery project, with optional search filtering
List all tables in a dataset with optional search filtering
Execute a BigQuery SQL query and return the results
Output schemas not documented. LLMs cannot plan downstream tool calls or extract return values without knowing the structure of responses. Each tool description states what it returns (e.g., 'returns results') but no formal output schema is visible in the code.
Parameter descriptions lack constraint detail. Parameters like 'top_k' (vector_search, generate_and_search) have no minimum/maximum bounds. 'distance_type' accepts 'COSINE, EUCLIDEAN, or DOT_PRODUCT' but this is mentioned only in the description, not as an enum constraint. Descriptions should state format, range, and allowed values explicitly.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 50 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 26 | - | v1 |
Search for tables across all datasets by name or column name
Perform semantic search using vector embeddings stored in BigQuery tables
No error handling guidance visible. The run_query tool description mentions 'must be a read-only SELECT or WITH query' but no documentation of what happens if a write query is passed, how the error is reported, or what the LLM should do. No recovery guides for common errors (table not found, permission denied, query timeout, quota exceeded).
Result limits not explicitly documented in tool descriptions. The code references DEFAULT_LIST_MAX_RESULTS (500) and DEFAULT_LIST_MAX_RESULTS_DETAILED (25) but tool descriptions do not state 'returns up to X results' or 'pagination not supported'. LLMs cannot reason about whether results are complete.
run_query and vector_search have compound concerns. run_query accepts dry_run and use_query_cache flags that are orthogonal to the core 'execute a query' action. vector_search and generate_and_search both apply filters and distance metrics in the same tool. Consider whether these should be separate tools or clarified in descriptions.
Parameter 'table_path' used in get_table, get_table_schema, vector_search, and generate_and_search accepts format 'dataset.table' or 'project.dataset.table' but this is not validated in visible code. If format is wrong, what error is returned? No error classification (retryable, user-fixable, fatal).