Single tool 'send_email' has a proper verb-noun name and comprehensive schema definition. However, the description in the actual tool registration (from list_tools()) differs significantly from the sampled version shown in the evaluation request. The schema visible in code includes proper types, descriptions, and defaults for all parameters. Critical issue: API key is exposed as a required parameter in the actual implementation shown in chatgpt_mcp_server.py (api_key field in inputSchema), which violates the secret-injection pattern. The description acknowledges stateless operation and no server-side storage, but the implementation contradicts this by requiring the user to provide the API key per-request as a parameter. This is a security anti-pattern. Parameter descriptions are generally good (80-150 chars), but the tool lacks output schema documentation and error handling guidance. The server claims 'no API key stored server-side' but forces the user to pass it as a tool parameter, which will be logged in agent traces.

• Move API key handling to server-side initialization via environment variable or secure header injection. Remove 'api_key' from the inputSchema. Document in server README that users must configure LEMON_API_KEY env var before deployment.
• Add explicit output schema to list_tools(): Document that send_email returns {success: bool, status_code: int, response: string, error?: string}. Enable LLM to detect failures and extract error messages.
• Enhance error responses with recovery guidance. E.g., if status_code=401, return error: 'Invalid or expired API key. Check that LEMON_API_KEY is configured correctly in server environment.' If status_code=422, include details from response body and suggest checking email format, sender verification, etc.
• Add enum constraint or pattern validation for 'to' parameter: describe as 'RFC 5322 compliant email address (e.g., user@example.com)'. Add example: not 'user_123' but a pattern description.
• Document the email confirmation UI workflow mentioned in description. Clarify: does ChatGPT's integration show a 'send email to X?' prompt, or is user consent handled elsewhere? If confirmation is required, document retry logic.
• Add 'variables' parameter description to clarify when template variables are used. E.g., 'Key-value pairs for Lemon Email template substitution. Example: {"name": "John", "verification_code": "ABC123"}. Only used if body contains template placeholders.'
• Reconcile schema discrepancy: the sampled inputSchema from evaluation request does not match the code in chatgpt_mcp_server.py. Ensure one authoritative definition.