An MCP server that provides tools to query Douyin (TikTok) video data analytics, including video lists, likes data, play statistics, and engagement metrics. Integrates with FastMCP and uses Alibaba Dashscope LLM for data analysis.
This server has critical definition quality gaps. All three tools are registered with `@mcp.tool()` decorators in `server/douyin_open_api_server.py`, so they are explicitly defined (not inferred). However, they suffer from severe naming, description, and parameter quality issues. Tool names use Chinese characters and lack action verbs, making them opaque to LLMs. Descriptions are minimal and lack guidance on when/why to use each tool or what data is returned. Parameter descriptions are extremely brief. No output schemas are documented. The server returns hardcoded mock data (acknowledged in a comment as a demo), which is a critical flaw for a production tool. Input validation is absent. Error handling is non-existent, there is no error classification, recovery guidance, or actionable error messages.
查询视频列表, 返回视频标题、转发数、评论数、点赞数、下载数、播放数、分享数
视频播放数据, 返回日期(date),播放数(play)
视频点赞数据, 返回日期(date),点赞数(like)
Tool names in Chinese without action verbs. LLMs cannot infer intent from non-English names and lack verb signals (get, list, create, search). '视频列表数据' is opaque; 'list_videos' would be immediately clear.
Descriptions are under 50 characters and lack context. '查询视频列表, 返回视频标题、转发数、评论数、点赉数、下载数、播放数、分享数' describes WHAT fields are returned but not WHEN to call this tool, what it does internally, or how it differs from similar tools.
Input parameter descriptions are too brief. 'User open ID for Douyin API' and 'Video item ID' lack format constraints, length limits, or validation rules. Parameters need descriptions of the form 'The project key (2-10 uppercase letters)' to guide LLM input.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 21 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 32 | - | v1 |
No output schemas documented. Tools return dicts with fields like 'data_list', 'chart_url', 'date', 'like', 'play' but LLMs have no schema definition to know what to expect. Per pattern, all tool outputs must document their schema.
Hardcoded mock data with no real API integration. The code comments state '这是一个demo的演示,对应tool的数据都是固定的' ('This is a demo; all tool data is fixed'). Tools return static results regardless of input, making them non-functional for any real use case. This is a showstopper for production deployment.
No error handling or validation. Tools do not validate open_id or item_id formats. No try/catch blocks, no error messages, no guidance for recovery. If an API call fails or parameters are invalid, the tool will crash or return misleading data.
Tool composition issues: tools 2 and 3 (点赉数据, 播放数据) require item_id but tool 1 (列表数据) does not return item_id in the visible output schema. This breaks the tool chain, an LLM cannot easily extract item_id from list results to pass to detail tools.
No pagination support. Tools return unbounded lists (e.g., video_list has a 4-item hardcoded array). Real video lists could have thousands of items. Per pattern, list tools must support limit/offset and return total counts.