MCP server for Volcano Engine (ByteDance) Doubao LLM models -- chat, text generation, embeddings
The server defines 4 tools with visible schemas and descriptions in src/main.rs. However, critical gaps limit quality: (1) descriptions are generic and under 100 chars, lacking WHEN-to-use and recovery guidance; (2) parameter descriptions are sparse or missing entirely; (3) no input validation or error recovery patterns; (4) output schemas not formally documented; (5) tool names lack action verb clarity (e.g., 'doubao_generate' vs 'generate_text_with_doubao'). Naming follows underscore convention but is vendor-centric rather than action-centric. The tools handle LLM/embedding API calls correctly but the interface is underdocumented for agent composition.
Have a multi-turn conversation with Doubao models. Supports conversation history.
Generate text embeddings using Doubao embedding models. Useful for semantic search and RAG.
Generate text using ByteDance Doubao models. Use for text completion, content generation, and single-turn tasks.
List available Doubao models from Volcano Engine
Tool names are vendor-qualified (doubao_*) rather than action-verb-first (generate_text, start_chat, embed_text, list_llm_models). Current names force LLMs to reason about vendor identity rather than intent.
All tool descriptions are under 100 characters and lack critical WHEN-to-use guidance. 'Generate text using ByteDance Doubao models' does not explain when to call generate_text vs chat, or what differentiation factors exist. Baseline average is 194 chars; descriptions should answer: What, When, and What's Returned.
Parameter descriptions are missing or minimal. 'model' param in all tools has a description like 'Model ID (e.g., doubao-pro-32k, ...)' but no guidance on valid choices beyond examples. The 'messages' param in doubao_chat is underdocumented, no validation of role values or content length.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 41 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 39 | - | v1 |
Output schemas are not formally documented in tool definitions. Tools return JSON with fields like 'text', 'model', 'usage', 'finish_reason' but these are inferred from code, not declared. Per pattern:tool, agents need documented return types to plan downstream calls.
No error recovery guidance. If API call fails (e.g., invalid model, quota exceeded, network timeout), the tool returns a generic error string. Per pattern:recovery-guide, errors must tell the LLM what to do next, 'Invalid model: use doubao_list_models() to see available models' vs 'API error: 400'.
Tool composition is incomplete. generate_text and chat both call /chat/completions endpoint but generate wraps single-turn as a chat message sequence. For agents to reliably select the right tool, descriptions must explain the trade-off: use generate for simple tasks, chat for multi-turn conversation.
No input validation constraints. 'max_tokens' param defaults to 2048 but accepts any u64, LLM could pass 1,000,000 causing API errors or timeout. Numeric parameters should declare bounds (e.g., 'max_tokens: 1-4096, default 2048'). Same for temperature (0.0-2.0).
Response fields not aligned with tool composition. doubao_chat returns 'response' while doubao_generate returns 'text', inconsistent naming makes multi-tool workflows confusing. Per mxe:response-field-naming, consistent output structure aids chaining.