Loading ToolBench
Collection of four specialized MCP servers for academic research: chemistry (PubChem/BindingDB), genes-ontologies (MyGene/UniProt/QuickGO/Reactome), structures (RCSB PDB/AlphaFold/Complex Portal), and variants (gnomAD/ClinVar/dbSNP)
13 tools with complete input schemas and descriptions. Naming follows verb_noun pattern (compound_by_name, similarity_search, binding_affinity, etc.). Descriptions are concise (50-150 chars) and action-oriented. However, output schemas are not documented in the source code, only input schemas are visible. Parameter descriptions are present but minimal (20-50 chars), lacking guidance on format, range, or dependencies. No error handling guidance visible. Tools are read-only (good for safety) but lack idempotent/destructive annotations. Composition is sound: each tool does one thing, and related tools (e.g., compound_by_name + similarity_search) chain naturally via SMILES/CID. Missing: output schema documentation, parameter constraints (enums, ranges), error recovery guidance.
Measured protein–ligand binding affinities for a target from BindingDB.
Look up ClinVar records (clinical significance) for a variant/gene term via NCBI.
Resolve a compound name to its PubChem identity + key properties.
Look up a gene across identifiers via MyGene.info.
List variants in a gene from gnomAD (via its GraphQL API).
Fetch the AlphaFold DB predicted-structure metadata for a UniProt accession.
Fetch experimental-structure metadata for a PDB entry from RCSB.
Output schemas not documented. Tools return JSON but LLMs cannot see field names, types, or structure. Agents must guess what fields to extract for downstream calls.
Parameter descriptions lack format/range guidance. E.g., 'threshold' in similarity_search says '0-100' but no enum or min/max constraint. 'limit' and 'retmax' lack bounds documentation.
No error handling guidance visible in tool definitions. Tools call external APIs (PubChem, BindingDB, UniProt, gnomAD, NCBI) but no recovery hints for timeouts, 404s, or rate limits.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
GO annotations for a gene product from EBI QuickGO.
Reactome pathways a protein participates in, by UniProt accession.
Search the EBI Complex Portal for curated macromolecular complexes.
2D Tanimoto similarity search in PubChem for compounds like a query SMILES.
Fetch a UniProtKB entry (function, features, cross-references).
Population allele frequency for a single variant from gnomAD.
No tool annotations (readOnlyHint, idempotentHint). All tools are read-only and idempotent, but this is not declared in the schema. LLMs cannot infer safety properties.
Parameter descriptions are minimal (20-50 chars). E.g., 'a chemical name or synonym' lacks guidance on what happens if multiple compounds match, or how to disambiguate.