Google BigQuery cloud-warehouse backend binding plugin for mcpg. Runs operator-fixed Standard-SQL statements against BigQuery over the REST jobs API (jobs.query) and returns rows as JSON, typed by the result schema. Auth is a GCP service-account key (JSON). Includes read-only guard, max_rows cap, and maximum_bytes_billed cost cap.
Two well-named tools with comprehensive input schemas and strong descriptions. Both tools follow verb_noun naming (bigquery_query, bigquery_dry_run) and include detailed parameter documentation with types, ranges, and constraints. Descriptions clearly state what each tool does and when to use it (10 - 1024 char range met). Read-only enforcement is explicit. However, output schemas are not documented in the visible code, the QueryOutcome and EstimateOutcome structs exist but their JSON field mappings are not declared in tool registration. Error handling is present but recovery guidance is minimal. No tool annotations (readOnlyHint, idempotentHint) visible in registration.
Run a BigQuery dry run (dryRun=true): validate and plan the query without executing it. Returns estimated bytes that would be processed, cache hit status, and result schema, but no rows. Incurs no charge.
Execute a Standard-SQL statement against BigQuery and return rows as JSON, typed by the result schema. Enforces read-only guard, respects max_rows cap and maximum_bytes_billed cost limit.
Output schemas not documented. QueryOutcome and EstimateOutcome structs exist but their JSON field mappings (rows, truncated, row_count, total_bytes_processed, cache_hit, schema) are not declared in tool registration. LLMs cannot plan downstream operations without knowing what fields to expect.
No tool annotations visible. Neither tool declares readOnlyHint, idempotentHint, or destructiveHint in registration. Both are read-only and idempotent, but this is not machine-readable for agent planning.
Error recovery guidance is minimal. Error messages from enforce_read_only() and build_client() are redacted but do not guide the LLM on next steps (e.g., 'Statement rejected as non-read-only. Allowed: SELECT, WITH, EXPLAIN. Rejected: INSERT, UPDATE, DELETE, CREATE, ALTER, DROP.').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
No pagination support visible. bigquery_query accepts max_rows but does not return a next_cursor or total_count. Large result sets cannot be iterated; truncated flag exists but no mechanism to fetch remaining rows.
Parameter 'use_legacy_sql' defaults to false but is not marked as optional in descriptions. Unclear if omitting it is safe or if it must always be specified.