Find a video.
Open a YouTube video, an X video, or a supported web video. Podcast feeds run through the audio path.
how it works
How Uoink works.
Click, corpus, AI, result. That loop is the product. The helper does the boring extraction work so your model gets context it can actually reason over.
Click Uoink.
The extension sends the URL to the local helper. The helper extracts transcript, frames, comments, metadata, and channel context.
Read the corpus.
A markdown file lands on disk and on your clipboard with timestamps, screenshot references, comments, and source links.
Ask your model.
Paste into Claude or ChatGPT, or ask your MCP agent to call Uoink directly.
Resolve the video context deficit.
Most AI workflows fail at video for a simple reason: the model gets a URL or a raw transcript and has to guess the rest. A transcript alone omits the slides, the code on screen, the audience corrections in the comments, and the channel context that tells you whether a video overperformed or just existed.
Uoink captures the richer shape. It turns the video into a local corpus: timestamped transcript, screenshots, comments, metadata, channel context, and optional classifier blocks. That corpus is readable by humans, pasteable into Claude and ChatGPT, and callable by MCP agents.
The workflow stays deliberately plain. Install the helper. Install the extension. Click Uoink. Paste, or let the agent skip the paste. The value is not that Uoink summarizes for you. The value is that Uoink gives the model the source material it was missing.
what lands in your clipboard
Anatomy of one uoink.
Same structure every time. Easy to read, easy to cite, easy for an agent to parse.
corpus markdownhuman readable
# Title and sourceURL, channel, upload date, duration, views## Transcripttimestamped segments with source anchors## Screenshotsinline budget + full folder on disk## Commentstop 50, likes, timestamps, disagreement signals## Classificationhook type, format, topic, performance tierwhy this beats a raw transcript
The model can connect what was said with what was shown and how the audience reacted. That matters for code walkthroughs, product teardowns, lectures, creator analysis, and podcast interviews.
Large payloads still need discipline, so the extension includes clipboard budgeting: screenshot caps, compression settings, and comment limits. Rich enough to reason over, small enough to paste.
three example workflows
Use it the way you already work.
Decode a hook.
Uoink a competitor video. Ask Claude or ChatGPT why the first 15 seconds work, where the promise is made, and what comments prove the audience noticed.
Audit a channel.
Run playlist mode on the last ten uploads. Compare hook type, format, comments, and performance patterns across the channel.
Pull a claim trail.
Uoink a long interview or podcast. Ask for every claim a guest made, then cite timestamps back to the original source.
send to your model, or skip the paste
Configure agents ->
Clipboard is the on-ramp. MCP is the upgrade.
Claude Desktop, Cursor, Cline, Continue, and ChatGPT Desktop can call Uoink directly through the local MCP server.
Where the corpus lives.
Uoink writes captures to a local folder you control and indexes the useful parts in SQLite FTS5. The exact path depends on platform, but the principle is stable: your corpus is ordinary files and a local index, not a proprietary cloud workspace.
On Windows, the helper lives under %LOCALAPPDATA%\Uoink and your captures land in your Uoink library folder. On macOS, the helper path is planned around ~/Library/Application Support/Uoink/. Topic folders and markdown sidecars make the archive portable.
That is the compounding asset. First capture solves a single problem. Fifty captures become a library. Five hundred captures become something your agent can search, cite, compare, and build from.