MCP server for teslamate-llm-bridge — exposes play tools to Claude Desktop / Cursor / Codex. Exposes 4 MCP tools to query TeslaMate driving data via the teslamate-llm-bridge, with complete image-generation flow support.
Server exposes 4 tools via fastmcp/STDIO with visible schemas and descriptions. However, critical gaps undermine quality: (1) descriptions contain example values and implementation details rather than LLM-optimized guidance; (2) parameter descriptions lack format/constraint details; (3) no error handling guidance or recovery paths; (4) tool names lack clear verb-noun structure (run_play, get_creative_prompt are acceptable but list_plays and generate_play_image are borderline); (5) no output schema documentation visible; (6) descriptions mix Chinese and English, creating ambiguity for LLM parsing. Baseline: 194 chars avg description, 72 chars avg param annotation, this server's descriptions are longer but less actionable.
用火山方舟 Seedream-4.0 文生图模型生成一张分享图,并以 base64 图片形式返回。调用前必须先走完整流程:1) run_play 拿 JSON 数据;2) get_creative_prompt 拿 plays/<name>/creative-prompt.md 模板;3) 把真实数据填入模板占位符,得到完整 prompt 后传入此工具。直接裸写 prompt(跳过模板步骤)会显著降低图片质量——模板内嵌了 5 轮迭代调优的 Seedream 专用措辞规律(9/10 评分验证)。需要设置环境变量 ARK_API_KEY(火山方舟 API Key);未设置时返回配置引导。中国大陆直连,无需 VPN,原生中文海报能力强,新用户 200 张免费。
读取指定玩法(play)的 creative-prompt.md 模板原文并返回。模板包含 v1(通用版,适用于 GPT Image / Qwen / Gemini)和 v2(Seedream 专用,适用于火山方舟 Seedream-4.0 / 豆包生图)两套 prompt,以及占位符填充说明和 Seedream 措辞规律。把 run_play 返回的真实数据填入模板占位符,再传给 generate_play_image 生成图片。找不到模板时返回中文操作引导。
List all available plays (analyses) loaded by the bridge engine. Returns each play's name, title, emoji, description, default time window in days, and whether a share card image is available. Call this first to discover which play names to pass to run_play / get_creative_prompt.
Run a play (analysis) for a given TeslaMate car and return a structured JSON result. If the car has fewer data points than the play's minimum sample threshold, scored=false is returned with sample count and minimum required. Use list_plays to see valid play names. The car_id is the integer primary key from TeslaMate's cars table (usually 1 for a single car).
Descriptions mix Chinese and English, creating parsing ambiguity for LLMs. get_creative_prompt and generate_play_image descriptions are primarily Chinese with embedded implementation details (e.g., 'Seedream-4.0', 'ARK_API_KEY', '9/10 评分验证'). LLMs cannot reliably extract intent from multilingual descriptions.
Parameter descriptions lack format constraints and validation rules. E.g., 'play_name' says 'kebab-case play name, e.g. driving-personality' but does not specify length, character set, or validation behavior. 'start_date' and 'end_date' say 'ISO 8601' but do not clarify timezone handling or what happens if start > end.
No output schema documentation. Tools return 'structured JSON result' (run_play) or 'base64 image' (generate_play_image) but exact field names, types, and nesting are not documented. LLMs cannot plan downstream calls or extract required fields without seeing the schema.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
No error handling guidance. Descriptions do not explain what happens on failure (e.g., 'car_id not found', 'play_name invalid', 'ARK_API_KEY missing'). LLMs receive no recovery path, they cannot self-correct or suggest alternatives.
Tool names lack consistent verb-noun structure. 'list_plays' and 'run_play' are clear, but 'get_creative_prompt' mixes 'get' (retrieval) with 'creative_prompt' (a template artifact), and 'generate_play_image' uses 'generate' (ambiguous, create? render? fetch?). Baseline: 90% of A+ tools start with action verbs; this server is inconsistent.