Loading ToolBench
Broad Institute gnomAD genomic variant database (GraphQL)
mcp-gnomad exposes 5 genomic query tools with basic schemas and minimal descriptions. All tools have input schemas with required parameters and type definitions, but descriptions are terse (10-60 chars, well below the 194-char baseline). No parameter descriptions beyond 2 tools. No output schemas documented. No error handling guidance. Tools lack verb prefixes (e.g., 'variant' should be 'get_variant'). No pagination, no result limits stated, no chaining IDs in responses. Security: no secrets exposed, but no audit logging or permission gates. Composition: tools are single-purpose but lack integration hints.
Gene info + variants. Accepts gene symbol (e.g. "BRCA1") or Ensembl gene id.
Variants in a genomic region (≤25kb recommended).
Gene / variant search (autocomplete).
Transcript + variants.
Variant by chr-pos-ref-alt (e.g. "1-55051215-G-A") or rsid.
Tool names lack action verbs. 'variant', 'gene', 'region', 'transcript' are nouns, not verbs. Should be 'get_variant', 'get_gene', 'query_region', 'get_transcript' to signal intent to LLMs.
Descriptions are too short (10-60 chars vs 194-char baseline). 'Variant by chr-pos-ref-alt...' lacks context on when to use this vs search. No guidance on expected output or error recovery.
No output schemas documented. LLMs cannot plan downstream calls or extract fields. GraphQL queries are visible in code but not exposed as tool documentation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
Parameter descriptions missing or minimal. 'dataset' param has a default note but no description of what it controls. 'chrom' has an example but no explanation of valid formats (chr1 vs 1 vs chrX).
No error handling guidance. If a variant is not found, the LLM receives a raw GraphQL error with no recovery hint (e.g., 'Try search() to find valid variants').