Skip to content

insodimension/youtube

v0.2.0MIT

Let your agent search YouTube, look at videos and channels, manage playlists, read or post comments, and get a finished film ready to upload.

GitHub

YouTube

What it is

A YouTube connector for your agent, plus three skills that take a finished film from a render to a published video: they package it, design its thumbnails, and upload it through YouTube Studio in your own signed-in browser.

What it contains

  • Nine native tools over the YouTube Data API v3: youtube_search, youtube_get_video, youtube_list_playlists, youtube_list_playlist_items, youtube_playlist_add, youtube_list_comments, youtube_post_comment, youtube_get_captions, youtube_channel_stats. Sign in with your own Google Cloud OAuth client (the one Drive and Calendar use works).
  • Skill youtube: how to use the nine tools, their quota costs and what the API will not do.
  • Skill youtube-package: scored title options, a description with chapters built from the edit's markers, sources and credits, tags, captions, the AI-use disclosure decision, Shorts and end-screen plans. It writes receipts/publish.json in the production folder and checks it with scripts/publish-check.mjs (plain node or bun, nothing to install): chapter rules, title and description limits, file hashes, a verdict for every generated or commentary asset, credits present in the description.
  • Skill youtube-thumbnail: three 1280×720 thumbnails as HyperFrames compositions, rendered with the Motion plugin, checked against the zones where YouTube draws over a thumbnail and at phone size with scripts/thumbnail-pack.mjs (needs ffmpeg).
  • Skill youtube-studio-upload: a runbook that drives Studio with the Browser plugin's browser_run, reads every field back, and records the video id and the read-back in receipts/publish.json.
  • Rule youtube-context: always-on guidance for the connector tools.

Who can use it

Off by default; turn it on in Plugins. The connector needs a Google sign-in through your own OAuth client with the YouTube Data API v3 enabled. The three film skills also need the Browser plugin (upload), the Motion plugin (thumbnails) and a production made in the Video space; the package and thumbnail skills run anywhere a production folder exists.

Limits and risks

  • The upload has not been run against a signed-in Studio. The way it finds elements, attaches a file, replaces text and reads it back was exercised on a stand-in page built like Studio. The labels come from YouTube's help pages and public automation projects. Studio changes its markup; the skill reads the page first and stops when a field will not read back.
  • It never handles your password. Sign-in, two-step verification and CAPTCHA are yours, in the Browser View.
  • The upload runs browser_run, which has full Node access on your machine and uses whichever saved browser profile you let it open, including its Google sign-in. The host asks before each run. Use a profile you are happy for the agent to act in, and watch in the Browser View.
  • Public needs your yes. A private or unlisted upload is the default and can be deleted; a public one is asked for by title and channel.
  • Disclosure is your decision. The skill sets out what YouTube says must carry the AI label and records your answer with reasons; it cannot tell whether a clip looks real, and commentary clips carry a fair-use judgement that is yours.
  • Thumbnails are 1280×720. YouTube recommends larger; the Motion plugin's stills are this size. Custom thumbnails need a verified channel. The overlay zones and the title and thumbnail statistics the skills quote are third-party measurements, labelled as such, and can change.
  • The API connector still cannot upload, comments are public and permanent-ish, and search costs 100 quota units a call.
  • Captions after upload go through Studio's Subtitles page and are the least certain step.

Build and test

bun run check runs the formatter and linter. The two scripts are plain .mjs with no dependencies; their tests are in test/ and run with bun test.