AlphaFoldMCP defines 3 tools with action-verb naming (get_* pattern), but suffers from moderate gaps in parameter documentation and schema completeness. All three tools have descriptions (194-254 chars, within baseline range), but parameter annotations are thin. Input schemas are present with basic types, but lack detailed constraints, enums, and validation guidance. Output schemas are not documented. The server targets a specialized scientific domain (protein structure prediction) with read-only operations, reducing risk, but parameter descriptions could be significantly richer to guide LLM reasoning about UniProt accessions, checksum formats, and annotation types.

• Document output schemas for all three tools. Include a sample response object with field names, types (string, number, array, object), and descriptions. For example, get_alphafold_prediction should document: returns an array of model objects, each containing pdb_id (string), confidence_score (number), aligned_confidence_measure (array of numbers), etc.
• Expand parameter descriptions with format constraints and validation rules. For qualifier: 'UniProt accession in the format AC (e.g., P12345, Q5VSL9). Accessions are 6 characters, case-sensitive uppercase alphanumeric + dash. If provided accession is not found in UniProt, call search_uniprot() first.' For annotation_type: 'Type of annotation to retrieve. Valid values: MUTAGEN (AlphaMissense variants), CONFLICT (sequence conflicts), NATURAL_VAR (natural variants), TOPOLOGY (membrane topology). Defaults to MUTAGEN.'
• Add error recovery guidance to all three tool descriptions. E.g., for get_alphafold_prediction: 'If the accession is not found, use search_uniprot() with a partial protein name. If checksum validation fails, re-fetch the sequence without the checksum parameter.'
• Clarify the output_json parameter. Document when to use JSON strings (if the response needs to be serialized/logged as-is) vs structured objects (for immediate field access). Consider removing the parameter entirely if structured objects are always preferred, simpler for LLMs.
• Add parameter inter-dependencies where applicable. For get_annotations: 'If annotation_type is MUTAGEN, the response includes per-residue missense scores. Other types may have different structures.' This helps LLMs anticipate response formats.

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

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