Skip to content
Back to Resources
GuideAdvanced

Give Your AI Eyes · Stop It Building Blind (agent-browser)

There's a silent problem with almost every AI that builds code: it works blind. It generates a screen, tells you "done, it works"… but it never OPENED it to check. It has no eyes. agent-browser — an open-source tool from Vercel — gives it exactly that: the ability to open your app in a real browser, look at it, click, fill out forms, confirm it actually works, and tell you what's broken. You install it in your agent (Claude Code, Cursor, Codex) and ask it to constantly VALIDATE what it builds, instead of guessing. And that's just the beginning: with eyes on the web, your AI can also browse sites, pull information, and automate repetitive tasks. In this guide you'll see how to install it, the prompt that makes it validate without you having to remind it, and the world of possibilities it opens up.

Jun 22, 202612 min
Who is this for?
For anyone building with a coding agent (Claude Code, Cursor, Codex) who wants it to stop guessing and actually check its work. It's the most advanced resource in the series (it needs a terminal agent), but the concept serves everyone: an AI that validates what it builds is an AI you can trust.

The problem · your AI builds BLIND

This happens to everyone who builds with AI, and almost nobody notices it: the AI generates a screen, a form, a flow… and tells you "done, it works." But did it open it to see? No. The AI writes the code and assumes it's fine — it has no eyes to check. It's like a cook who plates the dish without tasting it. Sometimes it comes out right; other times, the button does nothing, the login is broken, or the image doesn't load — and you're the one who finds out, in production, in front of a customer.

The pain this is born from · the "done" that wasn't done
That "it works" that was actually broken is one of the most frustrating moments of building with AI. You lose trust: every "done" is something YOU have to go verify by hand, opening the browser, clicking, checking. That's slow and exhausting. What if the AI did it itself, before telling you "done"?
A true story (ours)
This isn't theory. While building this very product, at first we'd ask the AI to go check what it had done — and it would answer, honestly, "I have no way to open the app to see it." It was building blind. Everything changed when we installed agent-browser: from that day on the AI started opening the app, navigating it, and checking what it built — and even validating these very guides live. The before and after is real, and it's exactly what this guide teaches you to set up.

What is agent-browser? · your AI's eyes

agent-browser is an open-source tool from Vercel that gives your AI agent the ability to drive a real browser: open your app, see what's on screen, click, type in fields, fill out forms, take screenshots, and read errors from the console. In plain terms: it gives it eyes and hands in the browser. No more guessing; now it checks.

Picture it like this · the cook who tastes the dish
An AI without agent-browser is a cook who cooks and plates without tasting — trusting it turned out fine. With agent-browser, it's the cook who tastes every dish before it goes out: opens it, looks at it, confirms it's good, and only then tells you "done." The difference between crossing your fingers and being sure.
Isn't this the same as webapp-testing?
They're similar, but agent-browser goes further. The webapp-testing skill (from the 5-skills resource) tests YOUR app. agent-browser does that and also browses any site on the web, pulls information, automates tasks, and comes with an MCP server and an encrypted credentials vault. It's a full browsing tool, not just a tester. The focus here is twofold: validating what you build + opening up the world of the web to your AI.

Benefit #1 · make it validate constantly (not blind)

This is the use that changes how you work: you ask your AI to, every time it builds something, open it and validate it itself before telling you it's done. It built a login → have it register with a test account and confirm it gets in. It made a form → have it fill it out and verify it submits. It changed a screen → have it open it and check that it looks right and there are no console errors. You go from "I think it works" to "I tested it and it works."

The habit · "don't tell me it's done without having seen it"
Make it a permanent rule for your agent: no task is considered finished without validating it in the browser. It's the same spirit as the C-A-R protocol, but for the interface: build → open and check → only then "done." An AI that checks itself makes far fewer errors that reach you.

Benefit #2 · it opens up a world of possibilities

Giving your AI eyes isn't just for validating your own work. With a browser at its disposal, it can do a lot more:

What your AI can do with eyes on the web
  • Validate your app end-to-end: register, navigate, test your key flow, and report what breaks.
  • Pull information from websites (prices, data, listings) and bring it back to you neatly organized.
  • Fill out repetitive forms and automate tedious browser tasks.
  • Research: open several pages, read them, and summarize what matters.
  • Reuse your session: if you're already logged into Chrome, it can work with that session (with your permission).

What makes it special (no jargon)

Unlike traditional automation tools (like Playwright), agent-browser is built for AI agents: instead of reading the whole page (which burns tons of tokens), it gives the AI a compact "map" of the elements it can interact with. That means cheaper and faster for your agent. It's open-source, from Vercel, with a permissive license.

vercel-labs/agent-browserREPO

agent-browser from Vercel — a browser automation CLI built for AI agents. It gives your agent (Claude Code, Cursor, Codex) eyes and hands in the browser: open, see, click, fill, extract, validate. It uses accessibility-tree snapshots (token-efficient) and comes with an MCP server. Open-source.

RustApache-2.0View on GitHub

How to install it (you need a coding agent)

You install it on your computer with a command, and the first time it downloads its browser. Then your AI agent uses it. (There are several ways — npm is the most common.)

Install agent-browser (npm)bash
npm install -g agent-browser
agent-browser install
If you'd rather, let your agent install it
As throughout the series: you don't have to wrestle with the terminal. Give your agent the repo link (github.com/vercel-labs/agent-browser) and tell it "install agent-browser and set it up so you can validate what you build in the browser." It handles it, and tells you if there's anything you need to do yourself. It's also on Homebrew (brew install agent-browser) and Cargo if you prefer.

Prompt 1 · Make it validate EVERYTHING it builds (the golden rule)

This is the prompt that changes how you work. Paste it into your agent so it stops building blind:

Paste it into your agent · validate, don't guesstexto
You have agent-browser installed. From now on, do NOT tell me something "works" without having checked it yourself in the browser with agent-browser.

Every time you build or change something in the interface:
1. Open my app (it runs at [your URL, e.g. http://localhost:3000]) with agent-browser.
2. Navigate to what you just built or changed.
3. Test it like a real user: click, fill in the fields, complete the key flow [describe your flow, e.g. "register → get into the dashboard"].
4. Check that it looks right and that there are NO errors in the browser console.
5. If something is broken or looks off, fix it and validate again, until it truly works.
6. Only then tell me "done," and tell me what you validated and what you saw.

Golden rule: no interface task is considered finished without validating it in the browser. Don't build blind.

Prompt 2 · Automate / pull info from the web

To take advantage of the other superpower — browsing and automating — this prompt asks your AI to use agent-browser for a specific web task:

Paste it into your agent · browse and extracttexto
Use agent-browser for this web task, step by step:

I want you to [describe your task, e.g.: "open this page [URL], pull the list of products with their prices, and bring them to me in a neatly organized table" / "go into my [service] dashboard and tell me the status of X" / "fill out this form with this data"].

1. Open the page and take a snapshot to see which elements you can interact with.
2. Do the task step by step, checking at each step that the correct element is there.
3. If I need to log in, tell me how to store my credentials securely (agent-browser has an encrypted vault; never put my passwords in plain text).
4. Give me the result neatly organized and tell me if anything couldn't be done.
Security · your credentials, encrypted
If the AI is going to log into sites with your session, NEVER paste your passwords into the chat. agent-browser comes with an encrypted credentials vault: you store the access by name and the AI uses it without ever seeing the password. And the session files (which carry tokens) go into .gitignore, never to GitHub. (This ties into the security guide in the series.)
Combine it with C-A-R
agent-browser is the perfect complement to the audit phase of the C-A-R protocol: when the AI audits its own work, it can now OPEN the app and truly confirm it works, not just re-read the code. Auditing with eyes is far more powerful than auditing blind.
In NeuralOS, validating is part of the flow
The "don't build blind" philosophy is at the heart of NeuralOS: what gets built gets checked, not assumed. The idea of an AI that validates itself before delivering is part of how we think about the product. If you want that rigor without setting up tools, that's the path.
The C-A-R protocol · audit with eyes, not blindagent-browser supercharges the audit phase: the AI checks in the browser, not just re-reading the code.Enterprise-grade security · protect your credentialsIf your AI browses with your session, store your credentials encrypted and never push the tokens to GitHub.
#agent-browser#automation#validation#claude-code