Stimmo demonstrates solid tool definition quality with comprehensive schemas, clear action-verb naming, and good parameter documentation. All 6 tools have explicit descriptions and well-structured input schemas with enum constraints for PropertyType, OmiCondition, FineCondition, etc. However, output schemas are not documented in the source code, the tool descriptions mention what is returned conceptually (e.g., 'Full Milan listing appraisal: geocode, OMI quote, amenity score, verdict') but no explicit response schema is visible. Error handling includes a custom ToolError class that returns structured {code, message} payloads, which is good practice. Tool naming follows verb_noun convention (estimate_property, lookup_omi_quote, geocode_milan_address, etc.), which is excellent. Parameter descriptions are present and detailed, including constraints like address format, enum values, and optional flags with defaults. The main weakness is the absence of visible output/response schemas, which prevents confident downstream chain planning. Additionally, no tool annotations (readOnlyHint, destructiveHint) are visible, though the tools are all READ_ONLY by nature.

• Document explicit output schemas for all tools. For estimate_property, show: {estimate_eur: float, min_eur: float, max_eur: float, geocode: {lat: float, lon: float}, zone: {code: string, name: string}, omi_quote: {min_eur_m2: float, max_eur_m2: float}, amenity_score: {transit: float, parks: float, shops: float}, verdict: string}.
• Add tool annotations in the MCP registration (estimated via server.py). For all 6 tools, set readOnlyHint=true since they are all queries with no side effects.
• Enrich error messages with recovery hints. E.g., ToolError('outside_milan', ...) should suggest 'Please provide an address within Milan city limits. Example: Piazza Duomo, Milano'.
• Add pagination support to omi_history if it can return multiple data points. If returning 8 semesters is always the limit, document this as 'Always returns the last 8 quarters of OMI data; no pagination needed.'
• Consider grouping related parameters. E.g., floor + total_floors could become a floor_info object {floor: int, total: int, has_lift: bool} to reduce parameter count and improve clarity.
• Document the caching behavior visible in _geocode_cached and _amenity_cached. Add a note to tool descriptions: 'Results are cached for 7 days (addresses) / 24 hours (amenities) to improve response time.'