Let your agentsee where humaneyes land.

Claude Code can read your screen. It can’t tell what people will notice. Heatshot predicts the human gaze, so your agent tests each design, compares versions and fixes what gets missed. Run the experiment before you publish.

Start for free

State of the art on design images

0.81

0agreement with real eye-tracking1.00

A real run on a sample page. No experiment, nothing published.

Reads anything people look at, one screen at a time

Web pages
App screens
Ads
Product images
Thumbnails
Packaging
Posters

Its maps agree with real viewers’ eye-tracking at 0.81, measured on 627 eye-tracked images that were never used to train the model: web pages, app screens, ads, product images and posters among them. It also reads thumbnails and packaging. How we check it

01 — What comes back

An answer your agent can act on.

Not a score to stare at. Where people look, what they miss, and which version is better.

POST /v1/analyze · MCP tool analyze_image

curl -s https://api.heatshot.ai/v1/analyze \
  -H "Authorization: Bearer $HEATSHOT_KEY" -H "content-type: application/json" \
  -d '{"image_url": "https://example.com/screen.png", "content_type": "web", "target_text": "Add to cart"}'

02 — Connect

Connected to your agent in a minute.

Heatshot is an MCP server and a JSON API. Add it once; your agent calls it whenever it draws a screen.

One command. Pick your agent and paste its setup.

Claude Code. Run this once in your terminal.

terminal

claude mcp add --transport http heatshot https://api.heatshot.ai/mcp/ \
  --header "Authorization: Bearer <your key>"

Or let the agent do it. Are you an AI agent? Read the setup page and connect yourself.

paste into your agent

Read and follow https://heatshot.ai/agents.md

Then ask it

Check this landing page with Heatshot and fix what people would miss.
Compare these two ad versions and tell me which one gets the button seen.
Does the price on this product page get noticed in the first three seconds?

03 — Accuracy

State of the art on design images.

Heatshot matches where real people look more closely than the best published attention model.

Match with where people looked, out of 1.00

Heatshot
0.81
Best published attention model
0.71
Typical commercial attention tool
0.62

Measured on 627 eye-tracked images that were never used to train the model.

04 — Features

Everything a study tells you, as data.

Attention map. Where people look, for the whole view and for the first one and three seconds.
Viewing order. What they see first, second, third, and each area's chance of being seen at all.
Focus, clarity, load. How concentrated attention is, how clearly the layout reads, how much it asks of the viewer.
Contrast. Text that cannot be read is named line by line, with the ratio it needs.
Benchmark. Web and app elements are compared with selected best designs in their category.
Findings. Eleven kinds, each with the measured value and the value that clears it.
Compare and rank. A control against up to ten variants, or versions of one design ranked by preference.
Pages, batches, reports. A live page screen by screen, batches of up to 100, results as an HTML or PDF report.

05 — How it is used

Build, look, fix, look again.

The loop your agent runs on its own, before a person reviews the screen.