Two tools with basic descriptions and partial schemas. fetch_paid_api has a clear description (145 chars) and complete input schema with types and descriptions for all three parameters (url, params, max_budget_sats). discover_agents also has a clear description (82 chars) and complete schema with two parameters (query, directory_url). However, neither tool documents output schema structure, responses are returned as raw strings without field definitions. No error handling guidance for LLMs (e.g., 'if payment fails, try X'). Tool names follow verb_noun pattern (fetch_, discover_) which is good. Parameters lack constraints (e.g., max_budget_sats unbounded, no regex for URL format). No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk classification present (READ_ONLY) but not integrated into tool metadata.

• Document output schema for both tools. E.g., fetch_paid_api returns {status: 'success'|'error', data?: object, error_message?: string, invoice_amount_sats?: number}. discover_agents returns {agents: [{name, description, price_sats, endpoint}], total_count: number}.
• Add error recovery guidance. Replace 'Error paying invoice via Alby: {e}' with 'Payment failed: {reason}. Retry with a higher max_budget_sats or contact the Alby service.' Include actionable next steps.
• Add parameter constraints: max_budget_sats (min=1, max=1000000), url (format=uri), directory_url (format=uri). Use JSON Schema format/pattern fields and repeat in descriptions.
• Implement dry-run or confirmation for fetch_paid_api. Add optional 'dry_run: bool' parameter. When true, return invoice details without paying. Require explicit confirmation before executing real payment.
• Add tool annotations to schema. Mark fetch_paid_api with destructiveHint=true (initiates payment). Mark both tools with readOnlyHint=true for discover_agents (no side effects).