Server has 4 well-structured tools with mostly clear naming and documented descriptions. All tools have input schemas with typed parameters. However, descriptions lack LLM-optimization guidance, output schemas are not formally documented, error messages are generic, and parameter descriptions are sparse. The server follows basic MCP patterns but misses several production-quality signal patterns for agent optimization.

• Document output schemas for all 4 tools. For GetSpendingAwardsByAgencyId, specify: returns {'results': [...], 'total_count': int, ...} with award field structure. For SearchByKeywords, document KeywordSearchResponse fields (award_id, recipient_name, award_amount, generated_unique_award_id, ...) and whether more_available flag is included.
• Add parameter validation: year should accept ISO 8601 YYYY format (4 digits), agency_id should reference GetAgencies output or accept known codes (e.g., enum or regex pattern). Document format in parameter description: 'year: fiscal year in YYYY format (e.g., 2024)'.
• Enhance error messages with actionable guidance: Instead of 'Received 404 from usaspending.gov', return '{"code": "NOT_FOUND", "message": "Agency id not found. Call GetAgencies() to list valid agency codes."}' or '{"code": "INVALID_YEAR", "message": "Year must be 4-digit YYYY (e.g., 2024). Got: abc"}'.
• Expand tool descriptions to 100-180 chars with explicit WHEN/HOW: GetAgencies: 'Get list of all US federal agencies with their tier codes and names. Call this first if you only have an agency name but need the code for GetSpendingAwardsByAgencyId.' (150 chars vs current 89).
• Add pagination guidance to SearchByKeywords: document that results are capped at 20, ordered by award amount descending. If agent needs all results, suggest calling with different year ranges or keywords to narrow scope.
• Expose response chain IDs explicitly. If SearchByKeywords returns generated_unique_award_id, confirm that agents can pass this directly to GetAwardInfoByGeneratedUniqueAwardId without transformation.