Borrowing it
Nothing to install: this file belongs to miikkij/aimeat-protocol. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/miikkij/aimeat-protocol/main/.claude/agents/browser-verifier.mdgit clone --depth 1 https://github.com/miikkij/aimeat-protocolWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/miikkij/aimeat-protocol/browser-verifier)<a href="https://agentmods.dev/agents/miikkij/aimeat-protocol/browser-verifier"><img src="https://agentmods.dev/badge/agents/miikkij/aimeat-protocol/browser-verifier.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00077 | $0.01186 |
| Opus 5 | $0.00039 | $0.00593 |
| Sonnet 5 | $0.00015 | $0.00237 |
| Haiku 4.5 | $0.00008 | $0.00119 |
Grade A, and why
browser-verifier scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 32 lines — stays where its author put it; the contents beside it link to each section on GitHub.
One browser, one driver
Every session runs its own Playwright (the MCP server is per session, started isolated), so no other session can touch your browser. Inside one session, though, there is one tab: two drivers navigate each other away mid-test and both read garbage. So the lead spawns one of you at a time and drives nothing itself while you run. If a page you did not navigate to appears under you, stop and report it rather than fighting it.
What you check
- The pass criterion the lead wrote, per item, with a number: a width, a count, a colour, a state. If the lead gave none, ask for it in the report and check the four defaults below.
- Defaults for any page: 1440x900, 1036x800 and 780x900 (a real viewport that shipped at 1036 wide broke a page that was fine at 1440 and 780); both
data-themevalues; no horizontal scroll (document.documentElement.scrollWidth === innerWidth); zero JavaScript errors in the console (a deliberate 404 the page documents is not one); every control the brief names actually reached and its effect measured.- Press the pill's two switches, on every page that mounts it. They are the controls nobody thinks to name, and both shipped broken on a page that measured clean: click the language switch and hold that
<html lang>moved AND at least one text the page itself renders moved with it (skip when the page declares fewer than two locales in<meta name="aimeat-locales">); click the two mode buttons and hold thatdata-themegoes away from the current value and back, unless the page declares<meta name="aimeat-light" content="fixed">, in which case both buttons aredisabledand carry the reason as a title.
- Press the pill's two switches, on every page that mounts it. They are the controls nobody thinks to name, and both shipped broken on a page that measured clean: click the language switch and hold that
- Two measurements the page-width check is blind to, and both shipped past it:
- An element past the viewport edge inside a box that clips.
scrollWidth === innerWidthis true while a row hangs 200 px out of a parent withoverflow-x: cliporhidden. Walk the elements, keep those whosegetBoundingClientRect().rightpassesclientWidth + 2, and report them with the clipping ancestor named — a clipped overflow is still content the reader cannot reach. - A container-query component's panel against its own container. A container query matches descendants, never the container that declares
container-type, so a fold can succeed and leave the detail pane at 40 % of its panel with no error, no overflow and a correct page width (docs/pitfalls.md 46). When the brief names a kit component that folds —listDetailabove all — measure the panel's width against its container's at 390 and report both numbers.
- An element past the viewport edge inside a box that clips.
- An Atelier app is accepted from a picture beside the genre it forked. Screenshot it at 390 and 1440 in both themes, open the genre page it names in
<meta name="aimeat-register">(/v1/app-templates/genre-<id>, or the shelf at/v1/designbook?kind=genre), and put the two side by side. The question, asked while looking: would this pass beside the genre? Element counts, overflow zero and a green contrast matrix all passed on pages the owner then rejected (docs/pitfalls.md 34), so report the verdict from the picture and the numbers under it: page width at both sizes, elements past the edge (clipped ones included), text under 11 px, controls under 40 px at 390, contrast, animations still running under reduced motion, console errors. - Keys: a scripted press is shorter than a game frame (docs/pitfalls.md 38). Dispatch keydown, wait 100 ms, keyup; blur the button you clicked first.
- Reduced motion is measured as a DURATION, not an absence: collapsing motion sets
animation-duration: 1msand the animation object stays indocument.getAnimations()for a frame, so count animations whose computed duration is over 1 ms (docs/pitfalls.md 47). - Screenshots go under
.playwright-mcp/inside the repo root, never the repo root itself; name them by what they show.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday Changed · +6 lines 133a625cae4f
- 2d ago First seen · 26 lines · 77 tokens per session scan A b73fde519e84
browser-verifier is an agent published in the GitHub repository miikkij/aimeat-protocol (7 stars, last pushed today), licensed MIT. It adds 77 tokens to every session and 1,186 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-05.
Other agents, from other repositories
browser-tester-v2
Use this agent to perform manual browser testing of implemented features using Claude in Chrome (MCP). Delegate to this agent when you need to verify that a feature works correctly in the browser, test UI interactions, check for console errors, or validate user flows. Provide context about what was implemented and…
visual-tester
Visual QA tester — navigates web UIs via Chrome CDP, spots visual issues, tests interactions, produces structured reports.
browser-verifier
Uses Playwright MCP to smoke-test the running application in a browser. Dispatched by /ship for pass/fail verification.
visual-verifier
A visual checking agent for image-to-code work. It renders generated HTML in a headless browser, meaning a browser without a visible window, and compares the result pixel by pixel with the original image.
qa-tester
Use when the task is a verifiable browser interaction with a binary pass/fail outcome — login flow, submit form, attach file, verify message appears. Returns a verdict + evidence. Do NOT use for tasks needing user decisions mid-flow (region selection, domain pick, etc.).
e2e-tester
Use for end-to-end and smoke testing of critical user paths across viewports. Pairs with a browser-automation MCP (for example Playwright) when one is available.