Four tools with complete descriptions and basic schemas. Tool names lack action verbs (resolve_commune, geocode_address, risques_immobilier, dpe_logement are noun-heavy or French-specific). All tools have parameter descriptions and type definitions. Output is unstructured text rather than structured JSON, forcing LLM parsing. Error handling returns plain text messages but lacks recovery guidance or categorization. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all being read-only operations. Descriptions are adequate (100-200 chars) but could be more concise for LLM optimization.

• Rename tools to verb-first pattern: resolve_commune → search_communes, risques_immobilier → get_property_risks, dpe_logement → search_energy_ratings. Improves LLM intent inference.
• Convert all tool outputs from plain text to structured JSON. Example: resolve_commune should return {communes: [{name: string, insee_code: string, department: string, region: string, coordinates: {lat, lon}, population: number}]} instead of formatted text.
• Add readOnlyHint: true to all four tools in FastMCP registration (@mcp.tool(readOnlyHint=True)).
• Enhance error messages with recovery guidance. Instead of 'Erreur : HTTP 404', return '{error: "Commune not found", suggestion: "Try a partial name or check spelling", retryable: true}'.
• Split risques_immobilier's overloaded parameter into two optional params (code_insee: string, adresse: string) with documentation: 'Provide either code_insee (5 digits) OR adresse (free-form). If both provided, code_insee takes precedence.'