The server defines 4 tools with explicitly visible schemas and descriptions in mcp_server.py. However, critical issues severely limit production readiness: (1) Most parameter descriptions are missing or minimal, only 'prompt' in submit_generation has a substantive description; 'query' and 'limit' in search_gallery have brief descriptions; others lack parameter documentation entirely. (2) Output schemas are not documented, responses are unstructured text or raw JSON dumps without schema declaration. (3) Error handling is minimal; tool errors are caught but responses lack guidance on recovery or next steps. (4) No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear semantic distinctions (query_status and list_workflows are read-only; submit_generation is destructive). (5) Tool names are reasonable but descriptions rely on Chinese language and cultural context unfamiliar to many LLMs, reducing cross-platform clarity. Naming follows verb_noun convention (query_, submit_, list_, search_), which is good. The schema syntax is correct JSON Schema, but parameter descriptions are sparse.

• Add explicit output schema documentation for each tool. E.g., query_status should declare: {"type": "object", "properties": {"online": {"type": "boolean"}, "running_tasks": {"type": "integer"}, "pending_tasks": {"type": "integer"}, "gpu_memory_used": {"type": "string"}}, "required": ["online", "running_tasks", "pending_tasks"]}
• Complete parameter descriptions. For list_workflows, clarify: 'Returns available ComfyUI workflows; no parameters required.' For search_gallery 'query' param, specify: 'Case-insensitive substring match against file names and paths. Empty string returns most recent files (sorted by modification time).'
• Add tool annotations to TOOLS array. Mark query_status and search_gallery with "readOnlyHint": true (no side effects, safe to retry). Mark submit_generation with "destructiveHint": true (initiates GPU work, may incur latency/resource cost).
• Translate descriptions to English or provide bilingual versions. Example: 'Submit a generation task to ComfyUI (occupies GPU). mode=image uses Flux text-to-image (auto-enhances Chinese to English); mode=video uses H3 text-to-video (~4.7 min per video). Call only when user explicitly requests generation; count ≤ 4.'
• Implement structured error responses. Instead of returning {"isError": true, "content": [{"type": "text", "text": "tool error: ..."}]}, add error code and recovery hint: {"isError": true, "content": [{"type": "text", "text": "Failed to connect to ComfyUI at {BASE}. Verify ComfyUI is running and COMFYAGENT_URL is correct. Retry after 10 seconds or check logs for details."}]}