One file, installed once, teaches your agent the whole YouTube transcript REST API — no MCP server to configure, just a free API key from the app.
Install the skill
mkdir -p ~/.claude/skills/youtube-transcripts
curl -fsSL https://bulktranscripts.co/skill.md \
-o ~/.claude/skills/youtube-transcripts/SKILL.mdAn agent skill is a Markdown file (SKILL.md) that teaches a coding agent such as Claude Code or Codex how to use a service. The BulkTranscripts skill documents the whole REST API, so the agent can fetch YouTube transcripts, search channels and monitor uploads with plain HTTP calls, no MCP client required. Install it by saving one file, then give the agent a free API key created at /app under MCP & API.
Everything below runs on the same engine as the web app.
Works for Claude Code today, and for Codex use ~/.codex/skills/ instead — any agent that reads the standard SKILL.md format works the same way. The raw file lives at /skill.md if you want to read it first (you should — it is short), and the source, examples, and issues live on GitHub: pratie/youtube-transcript-skill.
mkdir -p ~/.claude/skills/youtube-transcripts
curl -fsSL https://bulktranscripts.co/skill.md \
-o ~/.claude/skills/youtube-transcripts/SKILL.mdThe same skill is published on ClawHub, OpenClaw's skill registry, as @pratie/bulktranscripts-youtube (security audit: pass). One command installs it into an OpenClaw workspace; npx clawhub install bulktranscripts-youtube does the same for registry-managed skill folders.
openclaw skills install @pratie/bulktranscripts-youtubeFetch any video's transcript as clean text with timestamps and metadata. Search YouTube for videos, channels, or playlists. Search inside one channel's uploads. List a channel's videos (up to 1,000) or a playlist in order. Track a channel's newest uploads for free. Every capability is a documented curl call, so it works in any agent with shell access.
Skills are plain instructions — no process to run, no client-specific config, no protocol version to negotiate. Agent Skills are now a standardized, portable format across Claude Code, Codex, OpenClaw, and a growing list of agents: install once, carried in a folder, versioned in git if you like. If you prefer MCP, the same engine is a hosted server — see the YouTube MCP server page.
Sign in with Google at /app, open MCP & API → API keys, and create a key: free, 30 transcript credits, no card. Put it in BULKTRANSCRIPTS_API_KEY and the skill sends it as a Bearer token on every call. Buying a one-time credit pack tops up the same account, so the key never changes. Credits never expire, repeat reads from your own library are free, and the skill checks the balance in every response.
"Summarize this video and pull the key quotes." · "What has @TED published about AI? Fetch the three most relevant transcripts." · "Check if this channel posted anything this week and brief me." · "Get every transcript in this course playlist and turn it into study notes." The skill includes these playbooks, so the agent knows to list before fetching and to use the free latest-uploads endpoint for monitoring.
Anything that reads the standard SKILL.md format: Claude Code, OpenAI Codex, OpenClaw (via ClawHub), and compatible agents — plus any agent with shell access can simply follow the documented curl calls.
The skill file is free and redistributable. Usage rides the same allowance as everything else: 30 free transcript extractions, then one-time credit packs from $4.99. The new-uploads tracker is free forever.
Same engine, same credits, different transport. MCP suits clients with built-in server support (Claude connectors, Cursor, ChatGPT). The skill suits terminal agents and anywhere you'd rather have zero configuration.
Yes — it is a single markdown file at bulktranscripts.co/skill.md. Save it, edit the playbooks, commit it to your dotfiles. The skill file is freely redistributable; the API it calls is a hosted, proprietary service.
Sign in with Google, create a key, make the first call. No card.
Stuck on setup? Email hello@bulktranscripts.co with the app or code you are using and what you see. A person answers, usually within a day.
Updated