captions

YouTube Skills for AI Agents 🎬

Drop-in skills that let a coding or chat agent pull YouTube transcripts, run searches, walk channels and export whole playlists.

An Agent Skill is a folder of instructions an agent loads when the task matches. The 12 skills in this repo cover YouTube end to end. Each one names its endpoint, shows the exact request, lists every parameter with defaults and limits, and maps every error code to a recovery, so the agent makes the right call on the first try instead of guessing. No yt-dlp, no headless browser, no Google Cloud project. One HTTPS request per answer.

Works with Claude Code, Cursor, Antigravity, Cline, Codex, OpenClaw, Hermes Agent, and anything else that reads the Agent Skills format.

Made by TranscriptOut · API docs · MCP server

---

⚡ Quick start

npx skills add artemchuikin/youtube-skills --skill youtube-full

Then ask the agent something it previously had to refuse.

What does this talk actually claim? https://www.youtube.com/watch?v=dQw4w9WgXcQ

The first run also takes care of the API key, without leaving the chat. The agent asks for your email, a 6-digit code arrives in your inbox, you read it back, and a fresh sk_ key lands in your shell or agent config. The account behind it is free, starts with 100 credits and never asks for a card.

Hermes Agent

hermes skills install skills-sh/artemchuikin/youtube-skills/skills/youtube-full

All 12 skills at once

npx skills add artemchuikin/youtube-skills

Manual (git clone)

git clone https://github.com/artemchuikin/youtube-skills.git
cp -r youtube-skills/skills/youtube-full ~/.claude/skills/

OpenClaw (ClawdBot/Moltbot). Copy a folder from clawhub/ into your skills directory. The clawhub/ tree carries the same 12 skills with bash-style examples.

No terminal at all. Paste this into the agent and let it do the work.

Install the youtube skills from this GitHub repo: https://github.com/artemchuikin/youtube-skills
I want to get YouTube transcripts, search YouTube, and browse channel videos from my agent.
Set everything up for me.

Installing skills and creating the API key are both things these skills teach the agent to do.

---

💬 What you can ask

Every row is something you can say to the agent word for word. The skill turns it into the right API call.

---

🧰 Which skill to install

youtube-full is the right answer for most people. One skill, the whole surface. The narrower ones exist because a skill is context the agent has to carry, and a workflow that only ever needs transcripts should not pay for playlist documentation.

# one skill
npx skills add artemchuikin/youtube-skills --skill transcript

# several at once
npx skills add artemchuikin/youtube-skills --skill transcript --skill youtube-search

Beyond the core five, the rest are vocabulary variants of the same capabilities. People say "captions", "subtitles", "video to text" or "youtube api" for overlapping things, and an agent picks a skill by matching your words against skill descriptions, so each phrasing has a skill that answers to it.

---

💳 What calls cost

The unit is a credit, and one credit is one answered request.

A call that dies before reaching YouTube (bad parameter, rate limit, service busy) refunds itself automatically. A confirmed "this video has no captions" costs the credit, because the lookup really happened. The rate limit is 200 requests per minute per key on every plan.

Subscription volume is a slider from 1,000 to 100,000 credits per month in steps of 1,000, and the per-1,000 rate falls as the volume grows. Full pricing →

---

📦 Built for bulk work

Bulk export is one job, not a loop. The agent collects video IDs cheaply (ids_only=true returns up to 500 per page), submits them all in a single POST /v1/transcripts (up to 4,000 videos), and reads results page by page while the job is still running. Duplicates are dropped before billing, an Idempotency-Key makes retries free, the job paces itself inside the rate limit instead of tripping 429s, and any video the service fails to deliver is refunded on its own. A full channel archive that used to be an evening of scripting is three prompts.

🧠 Chunks sized for retrieval

Transcripts arrive pre-chunked if you ask. The segment parameter (20 to 5,000 characters) controls how the text is cut, and 500 to 1,500 gives retrieval-sized pieces with real context. Cut points come from per-word timestamps in YouTube's raw srv3 data, so each chunk's start is the moment the words are actually spoken rather