Give FrontierAgent the ability to read video. Transcripts, video and channel search, and handle resolution, as a drop-in FrontierAgent skill.
Video is a primary source most research agents cannot open. A conference talk, a lecture, an earnings call, a launch stream: the claim you need is in there, and your agent skips it. This skill turns any of them into quotable text with timestamps.
No yt-dlp, no headless browser, no binaries. One REST call, powered by
TranscriptAPI, which serves 15M+ transcripts a month.
Clone straight into FrontierAgent's skills directory:
git clone https://github.com/ZeroPointRepo/transcriptapi-frontieragent-skill.git plugins/skills/youtube-transcriptsAdd your key to FrontierAgent's .env:
echo 'TRANSCRIPT_API_KEY=sk_your_key_here' >> .envTurn skills on for the mode you run. FrontierAgent ships with skills: [], and a user profile at
~/.apodex/profiles/react.yaml overrides the built-in one:
skills: ["*"]Then run the terminal as usual:
uv run frontier-agent --mode reactFree tier: 100 credits on signup, no card. transcriptapi.com
| Ask your agent | What runs |
|---|---|
| "Summarize this talk: " | transcript, 1 credit |
| "Quote the part where they discuss pricing" | transcript with timestamps |
| "Find recent videos about agent frameworks" | video search, 1 credit |
| "What is TED's channel id?" | handle resolution, free |
A directory under plugins/skills/ holding a SKILL.md with YAML frontmatter. FrontierAgent
injects the name, description and path of every enabled skill into the system prompt, then the
model loads the full body with read_text only when the task actually calls for it. Two things
that are easy to get wrong:
- The profile role needs
read_textin its tools, or the model can see the skill and never open it. - Descriptions are truncated at 250 characters in the injected block. Write one that survives the cut.
Loaded against the real FrontierAgent skill loader (FileSystemSkillLoader +
SkillInjectionMiddleware) at commit ce4dd2d: discovered, frontmatter parsed, and injected into
the system prompt as an <available_skills> entry. Every endpoint in SKILL.md was run live
against the API before publishing.
SKILL.md the skill itself
scripts/transcript.sh a small wrapper that fails loudly when the key is missing
- youtube-skills: the same capability as Agent Skills for OpenClaw, Hermes, Claude Code, Cursor and Codex
- youtube-mcp: the MCP server, an Agent Plugin at the repo root
- awesome-frontieragent: every way to extend FrontierAgent
MIT licensed. Not affiliated with or endorsed by FrontierAgent or Apodex AI.