youtube-research-automation
Design agent workflows that pull YouTube transcripts, search videos and channels, and monitor new uploads without Google API quotas or OAuth setup
cosmicstack-labs · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add cosmicstack-labs/mercury-agent-skills --skill youtube-research-automation技能说明
name: youtube-research-automation description: 'Design agent workflows that pull YouTube transcripts, search videos and channels, and monitor new uploads without Google API quotas or OAuth setup' metadata: author: therohitdas version: 1.0.0 category: automation tags: - youtube - transcripts - video-search - channel-monitoring - research - agent-tools - openclaw - hermes-agent
YouTube Research Automation
Use this skill when an agent needs to read what was said in a video, find videos or channels on a topic, walk a channel or playlist, or watch a channel for new uploads, and you want that to run as a repeatable workflow rather than a manual copy and paste.
Recommended implementation:
- youtube-skills for any
runtime that supports the Agent Skills format. It ships twelve skills, and
youtube-fullcovers transcripts, video and channel search, channel browsing, and playlist extraction in one install. - youtube-mcp when the runtime speaks MCP instead of skills. Same tools, remote endpoint, API key or OAuth.
Both call TranscriptAPI, an independent REST API for YouTube content. It is not affiliated with YouTube or Google, and it is not the official YouTube Data API. That is the point of using it here: no Google Cloud project, no OAuth consent screen, and no daily quota unit accounting.
When to Use
Use this skill when the user asks for:
- The transcript, captions, or subtitles of a video, or a summary of one.
- Quotes, claims, timestamps, or numbers taken from a talk, lecture, or review.
- Videos or channels about a topic, without a browser in the loop.
- Every video on a channel, or a channel's recent uploads.
- Every video in a playlist, with titles and IDs.
- A recurring job that checks a channel and reports what is new.
- Research where video is fresher or more specific than written sources.
Do not use this skill for:
- Downloading or re-hosting video or audio files. This is a text and metadata workflow, and republishing someone's content is a licensing question, not a technical one.
- Uploading, commenting, account management, or anything that writes to YouTube.
- Videos with no captions. There is no transcript to fetch, so plan a fallback.
- Private or unlisted content the user does not own.
Core Model
Treat every YouTube workflow as three stages, and keep them separate:
- Resolve: turn whatever the user pasted, a URL, a video ID, an
@handle, or a playlist link, into a stable identifier. - Discover: search, list a channel, or list a playlist to get a candidate set.
- Read: fetch transcripts only for the candidates that survived filtering.
The reason to split stages 2 and 3 is cost. Discovery returns dozens of videos for one call. Reading is priced per video. An agent that fetches a transcript for every search hit burns the user's balance on videos nobody asked about.
Setup
One environment variable, one header convention.
export TRANSCRIPT_API_KEY="sk_..." # free key at https://transcriptapi.com, 100 credits, no card
curl "https://transcriptapi.com/api/v2/youtube/transcript?video_url=VIDEO_ID&format=text" \
-H "Authorization: Bearer $TRANSCRIPT_API_KEY" \
-H "User-Agent: YourAgent/1.0"
Send a real User-Agent. A missing or default agent string gets a 403 with
Cloudflare error 1010, and that failure looks like a bad key if you are not
expecting it.
Skills install:
npx skills add ZeroPointRepo/youtube-skills --skill youtube-full
MCP install, for runtimes that prefer a server:
claude mcp add --transport http transcript-api https://transcriptapi.com/mcp
Endpoints Worth Knowing
Base URL https://transcriptapi.com/api/v2.
| Endpoint | Use | Cost |
|---|---|---|
/youtube/transcript | Transcript of one video, format=json or text | 1 credit |
/youtube/search | Find videos or channels, type=video or channel | 1 credit |
/youtube/channel/resolve | @handle or URL to channel ID | free |
/youtube/channel/latest | 15 most recent uploads with view counts and timestamps | free |
/youtube/channel/videos | Full channel walk, 100 per page, continuation token | 1 credit per page |
/youtube/channel/search | Search inside one channel | 1 credit |
/youtube/playlist/videos | Playlist contents, paginated | 1 credit per page |
Channel endpoints accept an @handle, a channel URL, or a UC... ID directly,
so resolving first is optional. Full schema lives at
https://transcriptapi.com/openapi.json.
Workflow: Topic Research
# 1. Discover. One credit, up to 50 results.
curl "https://transcriptapi.com/api/v2/youtube/search?q=protein+folding+explained&type=video&limit=20" \
-H "Authorization: Bearer $TRANSCRIPT_API_KEY" -H "User-Agent: YourAgent/1.0"
# 2. Filter in the agent, not in the API. Drop by channel, age, duration, title.
# 3. Read only the survivors. One credit each.
curl "https://transcriptapi.com/api/v2/youtube/transcript?video_url=VIDEO_ID&format=text&include_timestamp=true&send_metadata=true" \
-H "Authorization: Bearer $TRANSCRIPT_API_KEY" -H "User-Agent: YourAgent/1.0"
Keep include_timestamp=true when the output will be quoted. A claim with a
timestamp can be checked by a human in ten seconds; a claim without one cannot.
Keep send_metadata=true so the title and channel travel with the text and the
summary can attribute correctly.
Workflow: Channel Monitoring
# Free. Poll this on a schedule and diff against what you stored last run.
curl "https://transcriptapi.com/api/v2/youtube/channel/latest?channel=@NASA" \
-H "Authorization: Bearer $TRANSCRIPT_API_KEY" -H "User-Agent: YourAgent/1.0"
Store the video IDs you have already handled. Fetch transcripts only for IDs that are new. Because the polling call is free, the cost of a monitor is e
阅读完整来源 ↗