Updated July 2026.
Screen recording feels like the obvious move when something's broken: hit record, narrate the bug, hand the file to your coding agent. It doesn't work — not because the idea is bad, but because Claude's API was never built to read video files. Anthropic's Messages API accepts images, not video containers (Source: Anthropic, 2026). The only route from a recording to something an agent can actually look at is turning it into a stack of still frames, and that route costs far more than it looks like it should. Here's exactly what happens along the way, what it costs, and what to send your agent instead.
Can Claude watch a video file?
No. Claude's API accepts image inputs — PNG, JPEG, GIF, WebP — not video containers like .mp4 or .webm. There's no native "watch this recording" call. Anything shaped like a video first has to be converted into a sequence of still frames before an agent can see any part of it at all.
Why doesn't my coding agent support .mp4 or .webm?
Coding agents like Claude Code and Cursor inherit whatever the underlying model API supports, and the model's vision input is images only. Native video support would mean reasoning over motion and audio streams directly — a materially different, heavier capability than today's per-image token pipeline (Source: Anthropic, 2026).
That's not a client-side limitation someone forgot to build. It's the shape of the API itself. Every tool that claims to let an agent "watch" a recording is doing the same thing under the hood: sampling frames, throwing away the parts a still image can't capture (motion, sound), and sending what's left as ordinary image tokens.
This is worth sitting with for a second, because it explains a lot of confusing behavior. If you've ever pasted a video link or a file path into a chat and had the agent respond as though it understood the recording, it didn't watch anything — it was almost certainly reasoning from a filename, a transcript you supplied separately, or a tool result, not from the pixels of the video itself.
What happens if you frame-dump a video into images?
Frame-dumping means extracting still frames from a recording at a fixed rate — commonly 1 frame per second — and sending each one as a separate image. A 30-second recording becomes 30 full-resolution screenshots, priced individually by Claude's image-token formula, stacked one after another into the same conversation.
Nothing about this process is smart. It doesn't know which second of the recording actually shows the bug and which second is just you scrolling past it. Every frame gets billed the same, whether it's the exact moment the button breaks or three seconds of an unchanged loading spinner.
Does a lower frame rate fix the token problem?
Not really. Dropping the sample rate lowers the token bill, but it doesn't fix the underlying issue: a lower frame rate is more likely to skip the exact moment the bug actually shows up. You're trading cost for coverage — cheaper, but blinder to the second that mattered.
There's no frame rate that reliably solves this, because a fixed-interval sampler has no idea which second matters — only the person who saw the bug happen does. Narrating the moment you see something break already carries that judgment call. A timer sampling frames every second, or every two, or every five, is just guessing. That's the deeper problem frame-dumping can't fix by dialing a slider: it treats every second of screen time as equally important, when almost none of it is.
How many tokens does a 30-second recording cost as frames?
About 80,730 tokens (calculated) — 30 frames at 1 fps, each a 1080p screenshot priced at roughly 2,691 tokens under Anthropic's published image-token formula. That's close to 40% of a 200,000-token context window gone (calculated) before your agent has read a single word of what you actually meant.
For the full arithmetic behind that per-image number — how it's calculated, what changes at different resolutions, and a real measured example — see the full token-cost breakdown.
How does a frame-dumped recording compare to just sending screenshots?
Frame-dumping is the most expensive way to get a screen recording in front of an agent — more expensive than taking a handful of screenshots by hand, and dramatically more expensive than a purpose-built review bundle. The table below lines up four approaches to the same review, labeled by whether the figure is measured on real hardware or calculated from Anthropic's published formula.
| Approach | Tokens | Type |
|---|---|---|
| 30s recording, frame-dumped at 1 fps | 80,730 | calculated |
| 4 screenshots covering the same page | 10,764 | calculated |
| Full Walkie bundle (transcript + pointed frames + REVIEW.md) | ~3,500 | measured |
| Spoken transcript alone, no images | ~150 | measured (range 100–200) |
The gap isn't subtle. Frame-dumping loses to plain screenshots because most of a 30-second recording is visually redundant — you're paying full image price for near-duplicate frames of a screen that hasn't changed. It loses even harder to anything that carries narration and intent instead of raw pixels, because words are nearly free next to images. Even the "cheap" screenshot column in that table still costs more than it needs to if the four screenshots overlap or include screen area that has nothing to do with the bug — the number improves further once you're choosing frames by what actually matters, not just capturing fewer of them.
What should I send my agent instead of a screen recording?
Skip the recording. Narrate the bug in your own words, take one or two targeted screenshots of exactly what's wrong, and state what you expected versus what happened. Words are nearly free — a spoken walkthrough runs about 150 tokens (measured) — and two well-chosen screenshots beat thirty redundant frames of the same mostly-static screen.
What to do instead of recording your screen for your agent:
- Narrate first, in plain language — describe the bug before you touch a screenshot tool. This forces you to name what's actually wrong instead of hoping pixels explain it.
- Point at what matters — crop or annotate so the agent's attention lands on the broken element, not the whole desktop.
- Take 1–2 targeted screenshots, not a reel of near-duplicate frames from a recording.
- Write the expected-vs-actual sentence yourself — don't make the agent infer intent from a pile of images.
For a step-by-step version of this workflow — and the tradeoffs of doing it manually versus with a tool built for it — see a practical workflow that avoids this entirely.
What do other screen-recording-for-AI tools actually do under the hood?
Every tool in this category — including Walkie — ultimately hands the agent images plus text, because that's the only visual input Claude's API accepts. The difference between tools isn't whether video gets sent raw; it's which frames get chosen and how much surrounding context — voice, pointer position, timing — rides along with them.
That distinction matters more than it sounds. A tool that samples frames blindly at a fixed interval is doing the same expensive, low-signal thing as manual frame-dumping, just automated — it inherits all the same problems this article walks through, faster and with a nicer UI wrapped around them. A tool that lets your voice and pointer decide which moments matter is solving a different problem: not "how do I get pixels to the model," but "how do I get the right pixels there, with enough context that the agent doesn't have to guess."
None of this makes Walkie the only option in this space, and it isn't. It's one answer to a problem every tool in this category has to solve the same way, because the constraint — images in, no video — comes from the model API, not from any single vendor's design choices.
The bottom line
Video-in, agent-reads-it-out was never on the table — Claude's API reads images, and a screen recording has to be broken down into a stack of them before an agent sees any of it. Frame-dumping does that the expensive way: 80,730 tokens (calculated) for 30 seconds at 1 fps, most of it redundant. Screenshots do it more cheaply. Narration plus a couple of targeted screenshots does it cheapest of all, because it replaces guesswork with a sentence.
None of this means recording your screen is pointless — it's a fine habit for your own reference. It just isn't the thing to hand your agent. Say what's wrong, point at it, and let the words carry the weight video can't.