Single tool 'convert_to_markdown' has a well-structured schema with comprehensive parameter coverage and clear descriptions. The tool definition is explicitly visible in src/md_server/mcp/tools.py and properly registered via FastMCP decorator in server.py. Naming is action-verb-based and descriptive. Description is substantial (~280 chars) and explains use cases, supported formats, and output modes. Parameter schema is complete with types, descriptions, enums, and sensible defaults. Output format is documented (markdown or json). However, output schema structure itself is not formally documented, and error handling lacks specific recovery guidance beyond what the implementation provides.

• Document the exact JSON structure returned when output_format='json'. Example: { title, description, content, metadata: { word_count, language, ... } }. This is critical so LLMs know which fields to extract.
• Add explicit error enumeration and recovery paths to the tool description: 'On unsupported_format error, check supported formats and retry. On timeout error, increase timeout parameter or set render_js to false. On invalid_url error, verify the URL is publicly accessible.'
• Add JSON Schema constraints to truncate_limit: 'minimum': 1, 'maximum': 100000 (for sections/paragraphs) or 1000000 (for chars/tokens). Document the bounds in the description.
• Formally declare url and file_content+filename as mutually exclusive in the schema using oneOf or by adding a note: 'Exactly one of url or (file_content + filename) must be provided. Passing both is an error.'
• Add a parameter 'idempotency_key' (optional string) so agents can mark calls they want to be idempotent. Store results by key to enable retry-safety.