This server demonstrates solid foundational quality with clear, action-verb tool names and well-structured parameters. All 5 tools are explicitly defined in src/main.rs with input schemas derived from serde/schemars. Descriptions are present and reasonably detailed (100-150 chars typical), with good parameter annotations. However, output schemas are not documented, LLMs cannot see what fields to expect from responses. Error handling is basic: the code uses anyhow::Result and McpError but responses shown do not include recovery guidance (e.g., 'model not found, try list_models() first'). The server follows tool composition well (each tool does one thing), and naming is clear and verb-forward. Security is properly handled: ANTHROPIC_API_KEY is injected via environment, never exposed as a parameter. The main gaps are: (1) missing output schema documentation, (2) no error recovery hints in tool descriptions, (3) no pagination documented for list_models despite it accepting a limit parameter.

• Document output schemas for all 5 tools. Add a new section to each tool's description or create a structured response definition (JSON Schema) that shows returned field names, types, and meanings. Example for ask_claude: 'Returns {"text": string (Claude's reply), "stop_reason": string (why it stopped), "usage": {"input_tokens": int, "output_tokens": int}}'.
• Enhance error responses with recovery guidance. Wrap API calls in custom error handlers that catch specific failures (model not found, invalid token count, API auth failure) and return human-readable messages like 'Model "claude-invalid" not found. Available models: claude-opus-5, claude-sonnet-4. Call list_models() to see all.' This enables LLM self-correction.
• Add pagination guidance to list_models. Clarify in the description: 'Returns up to [limit] models (default 20). If more models exist, a next_cursor field is included for follow-up calls. Use limit to page through results.' Include next_cursor and total_count in the actual response.
• Include example values in model parameter descriptions (e.g., 'Model id to use, e.g., "claude-opus-5". Defaults to the same model ask_claude uses.') to prevent hallucinated model names.
• Add tool annotations (readOnlyHint) to all tools. All five are read-only and safe to retry, annotate them with readOnlyHint=true so clients and agents know they have no side effects.
• Document max_tokens constraints in the chat tool. State 'Must be between 1 and [API limit, likely 4096]. Default is 1024.' This prevents LLMs from requesting absurdly large token counts.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry