Can Claude watch videos? Yes. Here's how, on your own machine

Claude reads text and images. It does not take video or audio as input, so a video link on its own gives it very little: at best the page around the video, the title and description, and nothing that was actually said or shown.

Claude Code is different because it can run programs on your computer. watch-for-me is an agent skill, a folder of instructions plus a script, that gives it the missing senses:

• Hearing. The audio is transcribed on your device by open speech models, with timestamps. No captions needed, no cloud.
• Sight. One frame per slide, scene or code change is pulled out and tiled into labelled contact sheets that Claude reads as images.
• Several at once. Up to 10 links per call, YouTube, Instagram Reels, TikTok, X or local files, processed in parallel.

So the honest answer to "can Claude watch YouTube videos?" is: not in the chat app, yes in Claude Code with this skill installed.

The default answer is a short summary, a timeline of what was said and shown, and any key text that appeared on screen. For YouTube, the timestamps come back as links that jump to that moment. Here is an abridged example of the shape, for a 1 hour talk:

**[1hr Talk] Intro to Large Language Models**, Andrej Karpathy, 59:48

Summary
- An LLM is two files: a parameters file and a short program that runs it
- Pretraining compresses a large slice of internet text into the parameters
- Fine-tuning on curated conversations turns that into an assistant
- The talk frames LLMs as the kernel of a new kind of operating system
- It ends on attacks: jailbreaks, prompt injection, data poisoning

Timeline
[00:20] LLM inference · slide: the two files of a 70B model
[04:17] Training · slide: GPUs, days and dollars for one training run
[14:14] Finetuning into an assistant · slide: a sample training conversation
[27:43] Tool use · demo: browser, calculator, Python, image generation
[42:15] LLM OS · diagram: the LLM as kernel, context window as RAM
[46:14] Jailbreaks · screenshots of roleplay and encoding tricks
…

Ask for a different shape with a flag: a 3-sentence TL;DR, a plain words explanation, a step-by-step checklist, the on-screen code as files, verbatim quotes, or the answer to one question with timestamp evidence. Or skip the summary entirely and ask for work: "watch this tutorial and add the same feature to my app".

watch-for-me ships as a standard agent skill, the same format as other Claude Code skills, so one repository covers every host. Pick yours:

Claude Code

Run these two commands inside Claude Code, one after the other:

/plugin marketplace add shafkathullah/watch-for-me

/plugin install watch-for-me@watch-for-me

Then use it as /watch-for-me. Autocomplete may show it as /watch-for-me:watch-for-me.

Any Agent Skills host

npx skills add shafkathullah/watch-for-me -g --skill watch-for-me

Codex

codex plugin marketplace add shafkathullah/watch-for-me

codex plugin add watch-for-me@watch-for-me

By hand

Clone the GitHub repository and copy its skills/watch-for-me/ folder to ~/.claude/skills/watch-for-me/.

Requirements

You also need uv and ffmpeg 5.1 or newer. Everything else installs itself.

curl -LsSf https://astral.sh/uv/install.sh | sh

brew install ffmpeg

The first run downloads the speech models once: about 2 GB for English, about 3.5 GB if you also watch non-English videos. To fetch them up front instead of on your first video, run this in Claude Code:

/watch-for-me --setup

Give it one link, several links, or a file, with an optional mode:

/watch-for-me https://www.youtube.com/watch?v=zjkBMFhNj_g

/watch-for-me https://youtu.be/aaa https://www.instagram.com/reel/bbb/ ~/Movies/demo.mp4

/watch-for-me https://youtu.be/xyz --ask "which GPU does he recommend?"

You can also just talk to your agent: "watch this and tell me the steps", "what does she say about pricing in these three videos?".

Modes combine, for example --tldr --quotes. Up to 10 links per call.

1. Download. yt-dlp fetches the audio and a 720p video stream in parallel, always the original-language audio, never an auto-dub. A 1 hour talk is about 50 MB.
2. Transcribe on your device. English goes to NVIDIA’s Parakeet model, other languages to Whisper large-v3-turbo, chosen minute by minute by a small language detector. On Apple Silicon it runs on the GPU; other machines use a CPU build.
3. Pick the keyframes. ffmpeg scene detection plus perceptual-hash deduplication keeps one frame per slide, scene or code change, then tiles them into labelled contact sheets. A 1 hour slide talk goes from 523 candidate frames to 73.
4. Your agent reads. Subagents read the sheets in parallel while transcription is still running, zoom into full-resolution frames for small text and code, and the main agent merges everything into one timeline.

Everything is cached per video, so a second question about the same video starts in under a second.

Privacy: audio and transcription never leave your device. The frames your agent reads go to your agent’s model provider, like anything else you show it. No telemetry. The only network calls are the video download and the one-time model downloads from Hugging Face.

The limit: watch-for-me sees keyframes, not motion. It catches every slide, scene change and line of code on screen, but it won’t judge a golf swing.

Measured on the v0.1 build in September 2026, on an M1 Pro MacBook with 16 GB of memory and a roughly 5 MB/s connection, models already downloaded:

Tokens: visuals cost about 200 tokens per minute of a slide talk and about 3,000 per minute of a fast-cut video, and most of that is spent in subagents rather than your main conversation. A transcript is about 13,000 tokens per hour of speech. Long transcripts are digested by subagents, so a 1 hour talk costs the main conversation about 13,000 tokens (2 hours: about 18,000).

Reels, TikToks and X videos often have no captions you can copy, which is why "instagram reel transcript" and "tiktok transcript" are such common searches. watch-for-me never relies on captions: it transcribes the audio itself, so a Reel, a TikTok or a video posted on X gets the same transcript and timeline as a YouTube talk.

• X. Public video posts work without a login. A post with several videos asks you to pick: --playlist 2 takes the first two.
• Instagram and TikTok. When a site asks for a login, add --cookies chrome (or your browser) to use the session you already have. Chrome on macOS asks for Keychain access; Safari needs Full Disk Access for your terminal.
• Local files. Pass a path instead of a link. A screen recording with no audio still gets a visual-only answer.

This is the one part of the page about our product. Deepmark saves everything you bookmark (links, X posts, Reels, YouTube) and lets you search it in plain English, down to what was said in a video.

Add --save to any watch-for-me call and the video’s link and title also go to your Deepmark library. Only the URL and title are sent, never the transcript or frames. To enable it, connect Deepmark to Claude Code once, then run /mcp to sign in:

claude mcp add -s user --transport http deepmark https://usedeepmark.com/api/mcp

In claude.ai, add the same URL under Settings, then Connectors. Deepmark is a paid service, $7 a month billed yearly or $10 paid monthly, with no free tier: without a plan the save is refused with a message saying so. Everything else on this page works without it.