Vllora MCP server exposes 2 read-only tools for trace querying. Tool naming follows verb-noun convention well (list_traces, search_traces). Both tools have substantive descriptions (95-180 chars). Input schemas are comprehensive with typed parameters and good descriptions. However, output schemas are completely undocumented, the server returns trace data but provides NO schema definition for the response structure. This is a critical gap: LLMs cannot plan downstream operations or extract specific fields when they don't know the response shape. Additionally, error handling is not visible in the code provided, and no tool annotations (readOnlyHint, etc.) are present despite these being read-only operations. Parameter descriptions are detailed and include enums and constraints where appropriate (e.g., operation_names, status, range_filter with predefined values).

• Document the output schema for both tools. Specify the structure of returned trace objects, including field names (e.g., trace_id, status, start_time, duration, spans, etc.), types (string, integer, object array), and descriptions. Example: '{"traces": [{"id": string, "status": "ok"|"error", "start_time_microseconds": integer, "duration_microseconds": integer, "operation_name": string, "spans": [...]}]}'
• Add tool annotations. Mark both tools with readOnlyHint=true to signal that they are safe to retry and have no side effects. This is now standard in the 2026-07-28 MCP spec.
• Enhance error handling documentation. In tool descriptions or a separate errors section, document possible failure modes and recovery guidance. Example: 'Invalid time_range: since must be before until. Check that ISO8601 timestamps are in the correct format (e.g., 2025-12-15T05:00:00Z).'
• Clarify optional parameter semantics. For list_traces, explicitly state: 'Omit run_ids, thread_ids, and operation_names to retrieve traces without filtering. If run_ids is an empty array [], no filtering is applied.'
• Document common label keys. In search_traces, provide examples or a reference for valid label keys to guide LLM usage. Example: 'Common labels include: {"agent": "name", "model": "gpt-4", "environment": "prod"}.'
• Add pagination guidance to tool descriptions. Both tools accept limit and offset; clarify typical patterns. Example: 'Default limit is 100. To retrieve all traces, make successive calls with increasing offsets until the response contains fewer results than the limit.'