TubeScout 🔭

Turn YouTube into a research engine for your AI agent. An MCP server (no API key) plus a skill pack that make Claude Code, Codex, and OpenCode search YouTube like a database, read transcripts at scale, and mine videos for evidence — claims, numbers, demand signals — instead of vibes.

Idea-engine tools scan Reddit and forums. YouTube is where founders show *receipts* — revenue dashboards, playbooks, real numbers on camera — and nothing mines it. TubeScout does.

Quickstart (60 seconds)

Claude Code

claude mcp add --scope user tubescout -- npx -y tubescout

Codex

codex mcp add tubescout -- npx -y tubescout

OpenCode — add to `~/.config/opencode/opencode.json` under `"mcp"`:

"tubescout": { "type": "local", "command": ["npx", "-y", "tubescout"], "enabled": true }

That's it — no API key, no config. Then ask your agent things like:

> *"Find the 5 most-viewed videos about n8n from the last month and summarize what people are struggling with."*

Easiest all-in-one (Claude Code): install as a plugin — MCP server + all 6 skills in two commands:

/plugin marketplace add not0lucky/tubescout
/plugin install tubescout@tubescout

Or install the skill pack manually (works for Claude Code, Codex, and OpenCode):

git clone https://github.com/not0lucky/tubescout && cd tubescout
./scripts/install-skills.sh   # installs into ~/.claude/skills, ~/.codex/skills, ~/.config/opencode/skills

Tools

Skills (the research methods)

All skills are context-aware: they read the conversation for what you're building, your stack, and videos already analyzed, and tailor verdicts to your actual leverage instead of giving generic advice.

See a real `/yt-breakdown` run on three "how I make $X/month" videos — including what survived the skeptic pass and what didn't.

How it works (honestly)

There's no magic here, and that's the point:

• youtubei.js talks to YouTube's internal InnerTube API — the same one the site uses. No key, no quota.
• Transcripts are YouTube's own captions, fetched through a fallback chain: the ANDROID-client timedtext track → the InnerTube transcript endpoint (known to 400 intermittently — retried with backoff) → local `yt-dlp` if you have it. Each response tells you which `source` served it.
• All analysis happens in *your* agent. The server ships data; the skills ship method.

Limitations

• Run it locally. YouTube aggressively rate-limits datacenter IPs — this is a local stdio server by design, not a hosted service.
• YouTube changes internals without notice; when it breaks, update (`npx` always pulls latest) and file an issue with the failing video ID.
• Videos with captions disabled can't be transcribed (rare; the error says so explicitly).
• Caption scraping lives in YouTube ToS gray area — fine for local research tooling, don't build a hosted paid product on it.

Development

npm install && npm run build
npm test          # unit tests (offline)
npm run test:live # live smoke tests against real videos — run before publishing
npm run inspect   # MCP Inspector against the built server

MIT — see LICENSE.

Built by Anir — I automate things. More at agramprojects.com.