uoink.video / v3.1

Uoink that shit.

One click under any video. Your AI gets the whole thing: transcript, screenshots, comments, channel context, podcasts too. Local. Yours.

Get Uoink See a live uoink

Open source / MIT / zero telemetry / 13 MCP tools / Windows now / Mac track in progress

karpathy-intro-to-llms.mdlocal corpus

# Intro to Large Language Models source: YouTube / channel: Andrej Karpathy saved: Desktop/Uoink/AI-and-ML/karpathy-intro-to-llms/ ## Metadata duration, views, upload date, channel context, source URL ## Transcript [00:47] hook classified as curiosity_gap + stakes ## Screenshots 12 frames on disk / 4 in clipboard budget ## Comments top 50 comments with themes and disagreements

one uoink / readable by Claude, ChatGPT, and agents

what you get per uoink

Why paste a transcript when you can feed the entire corpus?

Transcript-only tools flatten video into text. Uoink keeps the parts that make video useful: the words, the frames, the audience reaction, the channel context, and the source metadata.

Capture the source.

Transcript with timestamps, screenshots, description, title, thumbnail, channel context, and comments in one markdown file.

See the workflow ->

Keep the asset.

Every capture writes to disk and into a local SQLite index. Search it later, cite it later, move it into your own vault.

Browse features ->

Hand it to AI.

Paste into Claude or ChatGPT, send directly, or let an MCP agent call Uoink without touching the clipboard.

Open agent docs ->

local-first, by design

No account. No Uoink cloud. No telemetry.

The helper runs on localhost. Your corpus lands on your disk. Optional Hook Type, Comment Intelligence, and Entity Extraction calls use your own Anthropic key, not a Uoink proxy.

Read privacy ->

three audiences, one tool

Creators, researchers, and agent builders all hit the same button.

The on-ramp stays simple. The depth shows up after the corpus exists.

Creators.

Watch less. Read more. Audit hooks, pacing, screenshots, comments, and competitor channels without taking notes at 2x.

See creator workflow -> 02

Researchers.

Build a local library from videos and podcasts. Pull claims, guests, tools, products, and recurring entities across everything you saved.

Use it for podcasts -> 03

Agent developers.

Give Claude Desktop, Cursor, Cline, Continue, or ChatGPT Desktop 13 local tools for video capture and corpus search.

Wire up MCP ->

model agnostic

Uoink this video and compare the hook against my last ten saved competitor videos.

Calling uoink_video, then classify_hook, then search_uoinks. Your model reads the corpus. Uoink does not become the model.

mcp tool trace

uoink_video(url)
get_job_status(job_id)
get_uoink_corpus(corpus_id)
classify_hook(corpus_id)
find_mentions("Karpathy")

Read MCP setup -> Machine-readable MCP manifest ->

install in a minute

One helper. One extension. Then the U button.

Download the helper, install the extension, open a video, and click Uoink. The helper bundles what it needs so you do not install Python, yt-dlp, or ffmpeg yourself.

Windowslive path