Make Your Own Headless Browser (And Let Any AI Use It)

TLDR: I built my own headless browser — a real Chromium running on a small server — that stays logged in, turns messy web pages into clean Markdown, and that any of my AI tools can drive over a simple API. There are hosted ways to do this and they’re good. I wanted mine, mostly because once it’s wrapped as an MCP, every agent I make gets a browser for free.

👉 The build: github.com/alnutile/browser

---

Why I even wanted this

Here’s the wall I keep hitting.

I’ll be building some little AI thing — a research helper, a “go check this for me” agent — and it needs to read a web page. Easy, right? Except half the web doesn’t really exist until a browser runs it. You grab the page with a plain request and you get… an empty shell. All the good stuff loads later, with JavaScript. Or it’s behind a login, so you get a “please sign in” wall instead of the thing you wanted.

A person doesn’t have this problem. You open the page, it fills in, you’re already logged in, you read it. That’s all a browser doing its job.

So the idea was simple: give my AI that same browser. A real one, running on a server, that I can point at any page and say “load this like a human would and tell me what’s there.”

And I want to say up front — there’s more than one way to get here. There are hosted services that hand you a browser over an API, and they’re genuinely good; if you just need this to work tomorrow, reach for one. This is just the path that fit me this week, and I think the “why” is worth sharing even if you pick a different “how.”

What it actually is (in plain terms)

Strip away the jargon and it’s this:

Your AI  →  simple web call  →  a little server  →  a real Chromium browser  →  the live page

A small server sits there holding an actual browser open. Your AI doesn’t talk to websites directly — it talks to my little server, and the server drives the browser. Go to this page. Click that. Fill in this box. Now hand me back what’s on screen.

Three things make it genuinely useful instead of just a toy:

1. It stays logged in — even after a restart. This was the part I cared about most. A browser normally forgets everything when it closes. So I have it save its “profile” — the cookies, the logged-in state — onto a little bit of permanent disk (a Volume, in Railway terms). When the server restarts, it reads that profile back and it’s still logged in. No re-doing the login dance every time the thing reboots. The one trick that makes this work: shut the browser down gently on the way out, so it has a second to write that profile to disk before the process ends.

The little bit of permanent disk that keeps the login alive across restarts.

2. It gives back clean Markdown, not a mess. When you ask a page “what’s on you?”, you don’t want the ads, the nav bars, the cookie banners, the footer with forty links. So it runs the page through reader-mode (the same idea as the “reading view” button in your browser), strips it down to the actual article, and converts it to tidy Markdown. That’s the format AI models love. Point it at a cluttered page, get back the clean version.

3. There’s a lock on the door. Every request needs a secret token. No token, no browser. Because a browser-on-the-internet that anyone can drive is exactly the kind of thing you don’t leave open.

Two ways to drive it

This is the part I like, because it fits how I actually work.

Way one — spell out the steps. When I know exactly what I want, I send an ordered list: go here, type this in the search box, click the button, now give me the Markdown. It runs them in order and reports back on each one. Predictable, boring, reliable — perfect for a task I do the same way every time.

A list of steps in — per-step results out.

Way two — just say the goal. When I don’t want to think through every click, I send plain English: “Go collect every product on this site, and paginate through all the pages.” Behind the scenes an AI loops on the browser — look at the page, decide the next move, do it, check again — until the job’s done. Then it hands me the result and a transcript of everything it clicked, so I can see exactly what it did. That last part matters. I’m not blindly trusting it; I can audit the whole run.

Plain-English goal in — the final answer plus a transcript of every move it made.

The skill is in the browser repo as well Skill

Same little server. Same browser. Two gears — one for “I know the steps,” one for “you figure it out.”

The real reason: wrap it as an MCP

Okay, here’s the thing I’m actually excited about.

Everything above is just a web API. Useful, but it means each new AI tool I build has to know how to call it. Fine, but a little repetitive.

MCP fixes that. If you haven’t run into it yet, MCP is basically the standard plug that lets AI clients — Claude Desktop, Cursor, my own homegrown agents — discover a tool and use it without custom wiring each time. It’s the USB-C port of the AI world: one shape, and everything speaks it.

So I wrap this browser as one MCP tool. And now?

• My research agent can browse the live web.
• A little “check my dashboard and summarize it” bot can log in and read the thing.
• Claude Desktop itself can reach out and pull a real page.

…all sharing one browser that’s already logged in to the sites I care about. I build the browser once. Every AI I make after that just gets to “see the web” as a built-in ability. That’s the leverage. That’s why I bothered to make my own instead of renting one — a rented one is great, but this one snaps into everything else I’m building.

One tool in the list — and every agent can reach the same logged-in browser.

See the README.md

Was this hard? Honestly, no

I described what I wanted to Claude Code in Claude Desktop — a browser on a server, keeps its login, gives me clean Markdown, locked behind a token, drivable two ways — and worked through it with AI doing the heavy building. My job was the judgment calls: yes, put the profile on a Volume so it survives restarts. Yes, require a token. Give me a transcript so I can see what the agent did.

The pieces underneath are all off-the-shelf and battle-tested — Playwright to drive the browser, a small web server to take the calls, Railway to host it with that bit of permanent disk. None of it is exotic. The “hard” part was really just deciding what I wanted it to be, and then wiring good parts together.

The one thing I’d leave you with

You don’t need a browser company to give your AI eyes on the real web. A small server, a real browser, a place to save the login, and a lock on the door — that’s the whole recipe. Build it once, wrap it as an MCP, and every agent you make afterward inherits it.

There are other roads to this, and some of them are shorter. But I like owning this one. It’s mine, it stays logged in, and it plugs into everything.

The whole build is open if you want to read it or make your own:

👉 github.com/alnutile/browser