← Blog

Can Claude Analyze Videos? Three Ways That Work (2026)

Claude can't open a video file by itself. Anthropic's help center lists the files Claude.ai works with as documents (PDF, DOCX, CSV and the like) and images (JPEG, PNG, GIF and WebP), on a page updated July 23, 2026; video isn't on the list. What works is handing Claude something it can read: still frames and a transcript pulled from the video, or a connector that analyzes the video on its own servers and gives Claude the breakdown. Below are the three ways, a real breakdown of a clip we made, checked frame by frame, and the limits. A disclosure first: we make Riffkit, one of those connectors.

Three ways to get a video analyzed in Claude

Way Where it works What Claude gets Best for Watch out for
Frames and a transcript, with the Watch skill Claude Code (desktop Code sessions, VS Code, a terminal) and Codex; not Claude chat or Cowork Timestamped still frames, plus the captions or a transcript Finding out what a YouTube video, a talk or a screen recording says You install Python, FFmpeg and yt-dlp; its faster engine sends the video to Google's Gemini
Higgsfield MCP Claude, with a paid Higgsfield plan Higgsfield's read of the structure, shots and pacing, then new generations New clips in the style of a reference video Every generation spends Higgsfield credits; we haven't tested its analysis
Riffkit connector Claude and ChatGPT A timeline of what happens, the spoken lines, what's on screen and what each part does to the viewer Learning why a short video works, then making your version or changing your own clip Clips up to 45 seconds; on-screen words are described, not quoted

Pick by the job. To find out what a long video says, frames and a transcript are enough, and Watch pulls them for you. To get new footage in the style of a reference, Higgsfield's analysis feeds its own generators. To learn why a short video holds attention and make your own from it, or to change something in a clip you made, that is what Riffkit's breakdown is for.

What a breakdown looks like: our own 5-second clip

The clip is one we made with AI: a woman in a navy blazer sprints across an office plaza toward a low camera, under glass towers and an overcast sky, with the caption "me escaping 47 'where's my order' emails." It has music and no speech. We uploaded it on October 26 for a day-to-night Swap, and Riffkit analyzed it as the first step. In Claude, the connector's get_template tool reads back this same breakdown.

It came back in three parts:

  • A timeline. She bursts in from the lower right of the frame and fills two-thirds of it by 1.5 seconds; a cut at 3.33 seconds, quick stutter cuts at 3.46 and 3.58, a clenched fist pumped into the lower frame at 3.70, and a glance toward the right edge of the frame, checking behind her, at 4.15.
  • What's on screen. The woman, and the caption, listed as a meme hook that sets up running away from customer service inquiries.
  • Two parts, and what each does. 0 to 3.33 seconds is the attention part: panic and speed toward the lens take the viewer from passive scrolling to "amused recognition of hyper-specific customer support anxiety." 3.33 to 5.09 seconds is the payoff: the jump cuts and the look over her shoulder push the joke further.

What it got right. We checked it against frames from the file. She enters from the lower right a fraction of a second in, the fist comes up at about 3.7 seconds and the glance at about 4.15. FFmpeg's scene detection put the cuts at 3.333, 3.458 and 3.583 seconds, the times the breakdown gives. The blazer, the claw-clip bun, the gold chain and the towers all match.

Where it caught us out. In our day-to-night post we had called this clip one continuous shot. It isn't: those three jump cuts each skip ahead a few frames, and we hadn't noticed them by eye. We've corrected that post.

What it overstated. It calls the cut at 3.33 seconds "a hard cut to an eye-level close-up." The frames on either side of it are from the same low camera position, so it's a jump cut, not a new angle.

What it doesn't give you. The caption is described by its job, not quoted, so when you remake the video you write your own on-screen line. On a clip with speech, the breakdown lists each spoken line with its time; this one has none.

The analysis used no credits.

How to analyze a video in Claude with Riffkit

  1. Add Riffkit to Claude. Add a custom connector with the address https://mcp.riffkit.ai/mcp and sign in with your Riffkit account. How to make videos in ChatGPT and Claude has the setup for both.
  2. Give it the video. Paste a link to one TikTok video, or ask for an upload link for a clip on your device. A file attached to the chat doesn't reach Riffkit, so Claude gives you a Riffkit page to add it on; the link works for 90 minutes. Clips can be up to 45 seconds and 100 MB.
  3. Ask for the breakdown. On a Riffkit plan, Claude can analyze a TikTok link without making a video, for no credits. On a free account, the analysis runs as the first step of a remake. Either way, Claude reads the breakdown back and can walk you through it.
  4. Ask what to change, then for a price. Adapt makes a new video on the same structure for your product or character; Swap renders your clip again with one thing changed. Claude quotes the credits before anything renders.

Things to ask once it's connected:

Analyze this TikTok and tell me what each part does to the viewer: [link]

Here's my own clip. Break it down, then tell me what to change so the hook lands in the first second.

The other two ways, in short

Higgsfield MCP. Higgsfield's changelog entry for May 28, 2026 says Claude can take a reference video, analyze its structure, shots and pacing, and generate assets from that breakdown in the same chat, through Higgsfield's Supercomputer skill. The steps it gives: connect Higgsfield MCP in Claude, drop in a video, ask for a recreation. Its help center adds that the agent can't read files attached to the chat: files reach Higgsfield through an upload window the agent opens in the conversation. A paid Higgsfield plan is required, and generations use its credits. We haven't run it; how to use Higgsfield in Claude, Cursor and ChatGPT has the setup and the pricing.

The Watch skill. A free, open-source community skill (bradautomates/claude-video on GitHub) that turns a link or a local file into timestamped frames and a transcript for the agent. It runs in Claude Code and Codex; its README says it doesn't work in Claude chat or Cowork. The local engine needs Python 3.10 or later, FFmpeg and yt-dlp on your machine. The faster engine uses a free Google AI Studio key and uploads the video to Google's Files API, which deletes it after the answer, so keep private videos on the local engine.

What you can do for free on Riffkit

  • No card. New accounts start with signup credits.
  • The analysis uses no credits. On a free account it runs as part of a remake; the standalone analysis, with no video made, comes with a plan.
  • Free videos. Free accounts render at 480p, on Seedance 2.0 Fast or Seedance 2.0, with a small watermark, and a video made in more than one section gets its first section free.

Plans start at $15 a month for 3,000 credits.

Limits to know

  • Claude never sees the video itself. What it tells you rests on the frames, transcript or breakdown it was handed. Check the parts you'll act on against the video: ours called a jump cut a new angle.
  • Short clips only on Riffkit. Up to 45 seconds. For a long YouTube video, frames and a transcript are the better tool.
  • On-screen words are described, not quoted. Read the text off the video yourself when the exact words matter.
  • Use videos you have the right to use. Studying a video to learn why it works is one thing; reposting someone else's footage is another. Riffkit doesn't re-upload the source clip, and where the line sits covers the rest.

For the things a model tends to get wrong when it watches a viral video, see what an AI misses when it watches a viral video.

FAQ

Can Claude analyze videos?

Not by itself. Anthropic's help center lists documents and images as the files Claude.ai works with (page updated July 23, 2026), and video isn't one of them. Three routes work: give Claude still frames and a transcript pulled from the video (the community Watch skill does this in Claude Code and Codex); connect Higgsfield MCP, which analyzes a reference video's structure, shots and pacing and then generates new clips; or connect Riffkit, which returns a breakdown of a short video (a timeline, the spoken lines, what's on screen and what each part does to the viewer) and can then remake it.

Can Claude analyze video files like MP4?

Not as chat attachments: Claude.ai doesn't list video among the file types it accepts. In Claude Code, the community Watch skill reads a local file and turns it into frames and a transcript (it needs Python 3.10 or later, FFmpeg and yt-dlp). Through the Riffkit connector, Claude gives you an upload link instead: you add the clip, up to 45 seconds and 100 MB, on a Riffkit page, and Riffkit analyzes it on its servers and hands Claude the breakdown.

Can Claude analyze YouTube videos?

In Claude Code, yes, with the community Watch skill: it takes a YouTube link and produces timestamped frames plus the captions or a transcript (downloading from YouTube also needs a JavaScript runtime set up for yt-dlp). Its README says it doesn't work in Claude chat or Cowork. The Riffkit connector takes TikTok links and uploaded clips up to 45 seconds, not YouTube links, so for a long YouTube video, frames and a transcript are the better route.

Does Higgsfield MCP analyze videos in Claude?

Higgsfield says so. Its changelog entry for May 28, 2026 says that with Higgsfield MCP connected, Claude can take a reference video, analyze its structure, shots and pacing, and generate assets based on that breakdown in the same chat, through Higgsfield's Supercomputer skill. It needs a paid Higgsfield subscription, and generations use Higgsfield credits. Higgsfield's help center says files reach it through an upload window the agent opens in the chat, not as attachments.

How do I analyze a TikTok video with AI?

Paste the link into an assistant that has a video-analysis connector or skill. With the Riffkit connector in Claude or ChatGPT, a Riffkit subscriber can analyze a TikTok link without making a video, for no credits; on a free account the analysis runs as the first step of a remake. In Riffkit's test on a 5-second clip it made, the breakdown gave a second-by-second timeline and two parts (attention, then payoff) with what each does to the viewer, and placed three jump cuts at the times FFmpeg's scene detection found: 3.333, 3.458 and 3.583 seconds.

Is it free to analyze a video in Claude?

It depends on the route, since Claude needs help to analyze a video. The Watch skill is free and open source: its local engine runs on your computer, and its faster engine uses a free Google AI Studio key. Riffkit's analysis uses no credits: on a plan it runs on its own, and on a free account it runs as part of a remake, which free accounts make at 480p with their signup credits. Higgsfield MCP needs a paid Higgsfield plan.

Keep reading

How to Turn a Day Video into Night with AI

Change a video from day to night without reshooting. Real AI tests on our own clips: what turned to night, what stayed bright, and the steps.

How to Change the Color of Clothes in a Video with AI

Change the color of a shirt, hoodie or tank top in a video you already have, without reshooting. Real AI tests, what drifted, and the steps.

Formula Breakdown: The Walking Selfie Fit Check

Two shots in 5.3 seconds: a high-angle walking selfie, a hard cut to the feet, and one line about hard music on the way to a soft errand. The analysis card, and how to riff it.