Server has 21 tools with complete JSON schemas and French descriptions. Naming follows verb_noun convention (search_, create_, update_, delete_, list_, get_). Error handling and recovery guidance are absent. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Composition is reasonable but some tools could be consolidated (e.g., update_invoice_status + update_quote_status are near-duplicates).

• Add output schema documentation for all tools. Example: 'Returns: {invoice_id: string, number: string, status: InvoiceStatus, total_ttc: number, client_id: string, ...}' so LLMs know what fields to extract for chaining.
• Expand tool descriptions to 100-200 chars. Include WHEN to use (e.g., 'Call this after create_client to generate the first invoice') and what happens (e.g., 'Creates a draft invoice; use update_invoice_status to send').
• Add min/max bounds and format constraints to numeric/string parameters. E.g., 'quantity: number (min 0.01, max 999999)', 'postal_code: string (5 digits, French format)'.
• Implement confirmation/dry-run for destructive operations. Add a 'confirm' boolean parameter or a separate 'preview_delete_client' tool that shows what would be deleted before committing.
• Add tool annotations: mark delete_* with destructiveHint=true, get_*/list_* with readOnlyHint=true, and idempotent operations with idempotentHint=true per MCP spec.