The server defines 14 tools with explicit schemas and descriptions visible in src/deepresearch/server.py. Naming follows verb_noun convention (search_papers, fetch_paper_metadata, download_fulltext, etc.), which is appropriate. However, several critical gaps reduce quality: (1) Tool descriptions are generally adequate (100-150 chars) but lack guidance on WHEN to use each tool vs. alternatives, e.g., search_papers vs. search_apis vs. test_api_connector are not clearly differentiated. (2) Parameter descriptions are present but generic; many lack format/constraint details (e.g., 'paper_id' examples show format but no explanation of how to obtain one). (3) Output schemas are NOT documented anywhere, tools return unspecified structure, forcing LLMs to guess field names and types. (4) Error handling is absent from descriptions; no guidance on retryability, user-fixable errors, or recovery paths. (5) Two WRITE tools (store_to_drive, download_paper) lack confirmation/dry-run patterns and do not describe permission requirements. (6) Citation graph depth parameter accepts integers with no bounds (defaults to 1 but could accept unbounded values). This is typical of community MCP servers, functional definitions but missing production-grade polish around composition, error guidance, and output documentation.

• Add output schema documentation for all 14 tools. Specify field names, types, and whether arrays include pagination metadata (total_count, next_cursor). Example: 'Returns: { papers: [{id, title, authors, abstract, source}], total_count: integer }'
• Differentiate search_papers, search_apis, and test_api_connector in descriptions. E.g., 'Use search_papers for simple keyword queries across all sources; use search_apis for detailed source-specific configuration; use test_api_connector to diagnose API health or debug connection issues.'
• Add bounds to numeric parameters. Specify depth ∈ [1,5], max_results ∈ [1,100], max_citations ∈ [1,50]. Include rationale (e.g., 'limited to 100 to prevent context window exhaustion').
• Document paper_id format more explicitly. Add to descriptions: 'Format: SOURCE:ID (e.g. arxiv:2104.08935 or pubmed:12345678). To find a paper_id, first call search_papers() or fetch_paper_metadata() if you already know the source.'
• Add permission/scope declarations. For store_to_drive: 'Requires Google Drive write access. Will fail if folder_id is invalid or user lacks permissions, call test_api_connector(connector=drive) first to verify connectivity.'
• Implement error handling descriptions. For store_to_drive: 'On failure: returns error message indicating permission denial, folder not found, or network timeout. Retryable on timeout; non-retryable on permission error.'
• Add confirmation pattern for destructive operations. store_to_drive and download_paper should document optional 'dry_run' or 'confirm' parameter, or recommend wrapping in a verification step: 'Call with dry_run=true first to validate paper_id and folder_id before writing.'

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

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