The USASpending MCP server provides 2 tools with reasonable structure, but falls short of production quality in several key areas. Tool names follow the query_ verb pattern (good), descriptions are present and moderately detailed (average 180-220 chars, within baseline 194 range), and input schemas are fully visible with proper JSON Schema typing. However, critical gaps in output schema documentation, missing parameter descriptions in some fields, and no error handling guidance significantly limit agent reasoning. The decision_tree/router.py module shows parameter validation awareness but this logic is not exposed as tool behavior or error recovery patterns. Elicitation support (building missing param schemas) is a positive sign but underutilized, tools should raise and recover from missing required params automatically. Overall composition is clean: each tool does one thing (query agency data, query account data), but output structure is undocumented, forcing LLMs to guess what fields to expect and how to chain calls.

• Add explicit output schemas for both tools. For query_agency: document fields like agency_name, agency_code, agency_slug, budget_authority, budgets (nested), sub_agencies (array), etc. For query_accounts: show structure of federal accounts list vs. detail vs. TAS breakdown responses.
• Formalize breakdown enums in JSON Schema. query_agency: add enum constraint with all 7 values. query_accounts: list all valid breakdown types (not just mention 'object_class' default).
• Add mutual exclusivity constraints to query_accounts schema using oneOf/anyOf: one of (federal_account_id OR treasury_account_symbol) for detail modes, or neither for list mode.
• Expand parameter descriptions with validation rules and examples: agency_name: 'e.g., Department of Defense, DoD, or agency_code (format: 3-letter code)', fiscal_year: 'Federal fiscal year (1976-2026, defaults to current, Oct-Sep)', breakdown: 'Optional category for detailed breakdown. Valid values: [list]. Default returns agency overview.'
• Add error handling guidance in tool descriptions: 'If agency not found, the API returns a 404. Available agencies can be discovered via resource://agencies or by searching with partial names.' Similar for accounts.
• Enhance query_accounts description to clarify pagination: 'Returns paginated list of federal accounts (10 per page by default). Include total_count and next_page fields in response so agent can iterate if needed.'
• Document field mappings between query_agency response and downstream tools: 'Responses include agency_code needed for filtering in query_accounts when available.'