MCP server for interacting with Bilibili (B站) platform - enables account management, video operations, comments, user interactions, media downloads, and audio transcription using Whisper
bilibili-mcp has 14 tools with explicit schema definitions in internal/mcp/tools.go. Most tools have descriptions and input schemas present, but quality is uneven. Naming is action-verb-based and mostly clear (check_, list_, switch_, post_, reply_, get_, like_, coin_, favorite_, download_, follow_, whisper_audio_2_text, get_video_stream). Descriptions range from adequate (10-15 words) to verbose (40+ words), but many lack clarity on WHEN to use them or what they return. Parameters consistently have descriptions and type definitions, which is a strength. However, output schemas are entirely undocumented, no tool shows what fields are returned, their types, or structure. This forces LLMs to infer output shapes and guess at field names for chaining calls. Error handling is not visible in the provided code. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications. Several tools accept optional parameters with reasonable defaults, reducing required params. Composition is generally sound, single-responsibility tools with verb-noun naming. Critical issue: output schemas missing for all 14 tools; this alone caps the score significantly per the rubric ("Document the output schema").
检查B站登录状态
投币视频
智能下载B站视频媒体文件,优先下载包含音频的完整视频,仅在高清视频时使用音视频分离格式。支持实时进度显示和多种清晰度选择
收藏视频
关注用户
获取用户发布的视频列表
获取视频详细信息
No output schemas documented for any of the 14 tools. LLMs cannot know what fields to expect in responses, preventing reliable chaining and forcing inference of result structure. This violates the pattern:tool-composition rule.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications (READ_ONLY vs WRITE). Tools like post_comment, reply_comment, like_video, coin_video, favorite_video, download_media, and follow_user are destructive/write operations but carry no formal hint for the agent.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 56 | <=2025-11-25 | v2 |
| 2026-03-09 | C | 63 | 2025-03-26+ | v1 |
获取视频播放地址,直接返回可用的音频和视频流URL。只需提供视频ID即可,会自动获取第一个分P的播放地址
点赞视频
列出所有已登录的账号
发表文字评论到视频
回复评论
切换当前使用的账号
使用Whisper.cpp将音频文件转录为文字。支持多种音频格式,自动转换为最适合的格式进行识别。需要先运行 ./bilibili-whisper-init 进行初始化
Sparse or generic descriptions on read-only tools. Tools like check_login_status ('检查B站登录状态'), list_accounts ('列出所有已登录的账号'), and like_video lack depth on when/why to call them. No discovery guidance or context about prerequisites.
Descriptions in Chinese (e.g., '检查B站登录状态', '列出所有已登录的账号'). For international LLMs, non-English descriptions may be misunderstood or ignored. Tool descriptions should be in English or clearly indicate language context.
No pagination guidance for list_accounts and get_user_videos. get_user_videos accepts page/page_size parameters (good) but no documentation of result cap, total count, or next_cursor logic. Unbounded list responses risk context window exhaustion.
download_media and get_video_stream have complex parameter enums (quality values: 16=360P, 32=480P, ..., 127=8K; fnval values: 1=MP4, 16=DASH, ..., 4048=all). These are documented in descriptions but are not formal enum constraints in the schema. LLMs may hallucinate invalid values.
whisper_audio_2_text requires initialization via './bilibili-whisper-init' and model downloads. This prerequisite is mentioned in the description but there is no dependency hint or confirmation mechanism. Agents calling this without initialization will fail with unclear errors.
No error recovery guidance. Tools provide no documentation on what can go wrong, whether errors are retryable, or what the LLM should do next. For example, post_comment might fail if the user lacks permissions or if the video is deleted, but no guidance exists.