UniProt MCP presents well-structured tool definitions with consistent naming conventions and proper parameter documentation. All 5 tools follow verb_noun patterns (fetch, get, search, map) and have schema definitions. However, descriptions are generic, most hover around 50-80 chars without context on WHEN to use them or WHY they differ from similar tools. Parameter descriptions are present and mostly adequate, but error handling guidance is minimal. Output schemas are typed via Pydantic models (Entry, Sequence, SearchHit, MappingResult) and documented in code, but not surfaced in tool registration as JSON Schema. The tools are READ-ONLY (safe from destructive side effects), but lack explicit error recovery guidance for common failure modes (rate limits, invalid accessions, network timeouts). Overall, this is solidly competent but lacks the LLM-optimization depth needed for A-grade quality.

• Expand tool descriptions to 100-150 characters. Include WHEN to use: 'fetch_entry: Get full protein details by accession. Use when you need all data fields; for sequence-only calls, use get_sequence().'
• Add output JSON Schema to tool registration. Include 'Returns: { type: 'object', properties: {...} }' for each tool showing exact field names and types the LLM can expect.
• Convert map_ids from_db and to_db to enums. Document valid namespaces: ['UniProtKB', 'UniRef50', 'UniRef90', 'UniRef100', 'GO', 'ChEMBL', 'NCBI', ...] with descriptions.
• Remove fetch_entry version parameter entirely. Explicitly document that versioned access requires fetch_entry_flatfile(accession, version).
• Add enum constraint to fetch_entry_flatfile format: format: { type: 'string', enum: ['txt', 'fasta'], default: 'txt' }.
• Add error recovery guidance to each tool description: 'If accession is not found, try search_uniprot() first to locate the correct ID.'
• Remove ctx (MCP context) parameter from map_ids. Use MCP's built-in progress mechanism instead, or handle internally.
• Document search_uniprot result limits in description: 'Returns up to 500 results (capped automatically). Results are ranked by relevance.'
• Add pagination guidance to search_uniprot: 'For large result sets, use size parameter; implement client-side pagination if >500 total results needed.'

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry