ProxyLLM MCP has 9 tools with clear action verbs and reasonable descriptions (avg 180 chars). All tools have input schemas with type definitions and parameter descriptions. However, output schemas are not documented, responses are returned as raw JSON without field-level documentation. Error handling is minimal; most tools return API responses verbatim without recovery guidance. Parameter descriptions lack constraint details (ranges, formats, enums). No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications in metadata.

• Document output schemas for all tools. For proxyllm_list_models, specify: returns {data: [{id, object, created, owned_by, permission, root, parent}], object}. For proxyllm_account, specify: returns {id, object, created, plan, status, monthly_budget_usd, usage_usd}.
• Add recovery guidance to error responses. Instead of returning raw API errors, wrap them: 'No account token. Pass account_token parameter or set PROXYLLM_ACCOUNT_TOKEN env var. No account yet? Call proxyllm_signup first.'
• Add tool annotations to schema. Mark proxyllm_signup, proxyllm_verify_signup, proxyllm_create_routing_key with destructiveHint=true. Mark all READ_ONLY tools with readOnlyHint=true.
• Enhance parameter descriptions with constraints. For 'code': 'The 6-digit verification code sent to email (format: \d{6})'. For 'monthly_budget_usd': 'Monthly budget in USD (minimum: 0.01, maximum: 1000000)'. For 'mode': 'Routing strategy: fallback (try providers in order) or classifier (ML-based selection)'.
• Clarify token handling in all account-gated tools. Add to descriptions: 'If account_token is omitted, uses PROXYLLM_ACCOUNT_TOKEN environment variable. Required for authenticated endpoints.'