Agent's-Eye View
See any page the way a browser agent does: the screenshot and accessibility snapshot it reads, where it would get stuck, and how many tokens every look costs.
Created · Updated
What this tool does
Agent's-Eye View loads a page in a real Chrome browser and shows it the way a browser agent sees it: a screenshot of the first screen, and the text snapshot of the page's accessibility tree that agents read to decide what to click. Then it points out where an agent would get stuck: buttons with no name, clickable divs it can't recognize, pop-ups with no way out, and pages so large that every look burns thousands of tokens.
Why accessibility is agent optimization
Browser agents, the AI assistants that click through websites for people, work from some mix of screenshots and the accessibility tree, the same structure screen readers use. Playwright MCP, which many agents are built on, hands the model a text snapshot of that tree, and the snapshot in this tool is in exactly that format.
If a button has no name in the tree, the agent sees only "button" and has to guess. Fixing a page for agents fixes it for screen reader users too, and the other way around.
How it works
- Render: The page loads in headless Chrome at 1280×800, a common viewport for computer-use agents. The tool waits for the network to settle and for late pop-ups to appear, then takes a screenshot of the first screen.
- Snapshot: It captures the page's accessibility snapshot with Playwright's AI mode, element refs included, so every check runs on the tree an agent actually gets instead of a re-creation of it.
- Five groups of checks: Getting In (does the page load without a bot challenge, does anything cover it), Telling Things Apart (do controls and fields have names, do names repeat, are page regions marked), Taking Action (are clickable things real links and buttons, do menus say they open), Context Cost (tokens per look), and Your Goal when you give one.
- Your goal: Type a task like "add to cart" or "book a demo" and the tool looks for a link, button, or field whose name matches it, including common synonyms, then checks whether it sits on the first screen and whether a pop-up covers it.
- Fixes: Every problem lists the elements involved, a plain-language fix, and a prompt you can paste into Claude Code, Cursor, or another coding agent.
Reading the context cost
Agents re-read the page on most steps of a task, so the snapshot size is paid again and again. Token counts are estimates at about 3.5 characters per token. The screenshot figure uses Anthropic's published estimate for images (width × height ÷ 750). Raw HTML is shown for comparison: it is what an agent would read if it fetched the page instead of browsing it.
Limits
- One page, on load: The tool checks the page you enter as it first appears, without clicking anything. Menus, carts, and checkout steps behind an interaction are not opened.
- Heuristics: Pop-up, menu, and goal detection are heuristics tuned on real sites. A custom component can still slip past them or trip them.
- Blocked pages: When a bot challenge or error page comes back, the remaining checks are marked Not assessed instead of grading the challenge page. The tool detects challenges and never tries to solve them.