The server defines 7 tools with explicit schemas and descriptions in server/index.js. Most tools have clear, actionable descriptions (100-200 chars range), and all input schemas are properly declared with type definitions and additionalProperties guards. However, several critical gaps prevent a higher score: (1) output schemas are entirely undocumented, tool descriptions state what they call (GET /v2/domains, POST /v2/domains/{domain}/domainStats:query) but do not describe the response structure, fields, or which IDs/references are included for chaining; (2) error handling is minimal, responses are formatted with HTTP status labels but lack recovery guidance (e.g., 'run gpt_authenticate' is mentioned for auth issues, but other 400/403/404 responses offer generic instructions); (3) several tool descriptions embed example values (e.g., 'mail.example.com', 'example.com') which LLMs may reuse literally; (4) the three auth tools (gpt_authenticate, gpt_auth_status, gpt_sign_out) have empty input schemas and vague parameter documentation is absent, these are acceptable given they take no parameters, but descriptions could reference prerequisites more explicitly; (5) query_domain_stats has a complex metrics array parameter with anyOf union types, but the description mixes prose explanations of metric names and filter requirements rather than documenting the response format or pagination.

• Document output schemas for all query tools. For list_domains, specify: { domains: [{ domain: string, displayName?: string, createTime?: string, state?: string }], nextPageToken?: string, total?: number }. For get_compliance_status and query_domain_stats, provide example response structures showing all returned fields and their types.
• Enhance error messages with recovery hints. For 404 errors, suggest: 'Domain not found. Call list_domains() to see registered domains.' For 403 errors on missing scopes, guide: 'Insufficient permissions. Re-run gpt_authenticate to request broader scopes.'
• Remove embedded example values from descriptions and replace with formal constraints. Instead of 'Fully-qualified domain name, e.g. example.com', use 'Fully-qualified domain name (format: valid DNS domain, e.g. example.com or mail.example.com)' or document via JSON Schema pattern.
• Document the response schema for query_domain_stats explicitly. State: 'Returns an array of metric rows, each with { date: string (YYYY-MM-DD), metricName: string, metricValues: [{ value: number }] } for DAILY granularity, or { metricName: string, metricValues: [{ value: number }] } for OVERALL.'
• Clarify pagination in list_domains. Add to description: 'Response includes a nextPageToken field for retrieving the next page. Pagination is optional; omit pageToken to retrieve the first page.' Specify max result count (stated as 'default 10, max 200' in parameters) in the tool description as well.
• Add tool annotations to the schema. For gpt_sign_out, add "destructiveHint": true. For gpt_authenticate and gpt_auth_status, add "idempotentHint": true (auth status checks can be safely retried). These are optional MCP protocol features but improve agent behavior.