| name | video-transcript |
| description | Transcribe online videos (Bilibili/B站, YouTube, Vimeo, Twitch, and any yt-dlp supported platform) to text so agents can learn from video content. Use whenever the user provides a video link (bilibili.com, b23.tv, youtube.com, youtu.be, vimeo.com, twitch.tv, etc.), asks to transcribe a video, or wants to extract/summarize content from a video. Supports existing subtitles (fast) and AI-generated transcription via local Whisper (slower but always available). |
Video Transcription
Transcribes online videos to text using the video-toolkit_* native tools (backed by the video-toolkit MCP server configured in opencode.jsonc). Works with any platform supported by yt-dlp — Bilibili (B站), YouTube, Vimeo, Twitch, and hundreds more.
Workflow
Step 1: Resolve short links
Short links like b23.tv/XXXXX or youtu.be/XXXXX must be resolved first:
curl -sI "https://b23.tv/XXXXX" | grep -i location
Step 2: Try existing subtitles first
Call the native tool video-toolkit_get-transcript:
video-toolkit_get-transcript
Args: url="<video URL>", lang="zh"
- If the video has subtitles → instant result
- If no subtitles → fails, proceed to Step 3
Step 3: Generate subtitles with Whisper
Call the native tool video-toolkit_generate-subtitles:
video-toolkit_generate-subtitles
Args: url="<video URL>", language="zh"
Downloads audio, converts to 16kHz mono, transcribes with local Whisper (small model).
Timing (CPU, ~4 cores):
| Video length | Transcription time |
|---|
| 5 min | ~1 min |
| 15 min | ~3 min |
| 30 min | ~6.5 min |
| 1 hour | ~13 min |
Step 4: Use the transcript
Read the returned text and answer the user's question. Timestamps included — cite specific moments if needed.
Step 5: Check available languages
video-toolkit_list-transcript-languages
Args: url="<video URL>"
Supported Platforms
Anything yt-dlp supports:
Bilibili (bilibili.com, b23.tv) · YouTube (youtube.com, youtu.be) · Vimeo · Twitch · Podcast feeds · Direct media URLs · And hundreds more
Configuration
The video-toolkit MCP server is configured in opencode.jsonc and its tools are exposed as native video-toolkit_* tools (no skill_mcp invocation needed):
"video-toolkit": {
"type": "local",
"command": ["node", "/path/to/video-toolkit-mcp/dist/index.js"],
"enabled": true,
"timeout": 600000,
"environment": {
"YT_DLP_PATH": "/home/ubuntu/.local/bin/yt-dlp",
"FFMPEG_PATH": "/home/ubuntu/.local/bin/ffmpeg",
"FFPROBE_PATH": "/home/ubuntu/.local/bin/ffprobe",
"TRANSCRIPT_MCP_WHISPER_ENGINE": "local",
"WHISPER_BINARY_PATH": "/home/ubuntu/.local/bin/whisper",
"WHISPER_MODEL_PATH": "small"
}
}
The local whisper wrapper is bundled at scripts/whisper. Copy to ~/.local/bin/whisper and chmod +x.
Dependencies
yt-dlp — downloads video/audioffmpeg — audio format conversionfaster-whisper (pip install faster-whisper) — local transcriptionwhisper wrapper script at ~/.local/bin/whisper
Limitations
- Some platforms require authentication for subtitles (e.g. B站).
generate-subtitles with local Whisper always works as fallback.
- Transcription speed: CPU-bound, ~4-5x realtime with small model.
- Accuracy: Good for clear speech; noise or heavy accents reduce quality.
- Long videos (>1h) may exceed context windows. Process in chunks.
