A browser
your AI
can drive.

Before we start

Ever tried asking ChatGPT or Claude to go read a web page? Then you know the frustration. Sometimes it manages, sometimes it tells you "Sorry, I can't access that URL." Or worse: it hallucinates content it never saw, tells you things that don't exist, quotes made-up numbers. You click the link to check, and you realize the page is about a completely different topic.

Dev-browser fixes that. It's a windowless browser (Chromium behind the scenes, like Chrome without the ads) that your AI drives by writing a tiny script. You never open Chrome yourself. You say what you want, Claude writes the script, it runs in an invisible Chrome, and you get back whatever you're after · the text of an article, a precise figure in a table, a screenshot of a page, the render of a page you just edited. Under the hood, dev-browser relies on the same machinery as Playwright and Puppeteer (the two reference libraries for driving a browser), but you never touch their API: Claude does it for you.

I installed it on my Mac two months ago. Today, I use it almost every day: checking that a page is properly deployed, scraping a site Claude doesn't know, taking a screenshot to sign off a design, testing my site's signup form without opening Chrome by hand, capturing a competitor's article before it changes.

The lightbulb moment for me was when I realized dev-browser isn't a dev tool. It's a tool for a curious founder who wants their AI to actually do what they ask. You're no longer fighting with "I can't access," you're no longer copy-pasting content by hand, you're no longer opening ten Chrome tabs to check ten pages. You delegate, and your AI actually does it.

Here are the main alternatives at a glance, to place dev-browser in the ecosystem.

The concept · a browser you never see

The real problem it solves

Claude knows a huge number of pages thanks to its training. But there are two limits:

1. It doesn't know recent pages (created after its cutoff, meaning the date its training stopped).
2. It can't interact with a page: click, fill in, wait for a JS script to fire, scroll.

Dev-browser fixes both. It actually opens a browser (Chromium behind the scenes, like Chrome without the ads) with no graphical window. Claude sends it orders in JavaScript: "go to this page, wait 3 seconds, grab the text of the h1, click this button."

How it works, no jargon

Picture a robot that opens Chrome for you and does exactly what you tell it. That's dev-browser. Except the robot is your computer, and the Chrome window is one you never see. It all happens behind the scenes.

Concretely, here's what happens when you ask Claude "go check if my page is live":

1. Claude writes a tiny script (a 5-10 line mini-program, in JavaScript, the language of web pages).
2. The terminal launches Chrome in the background · what's called "headless" (no head, no visible window). It opens, but on a screen no one sees.
3. Dev-browser intercepts the script and runs it in that invisible Chrome: go to the URL, wait for the page to load, read the content, click a button, whatever you want.
4. The result comes back in your terminal · the text it grabbed, the screenshot image, the extracted data. Claude reads the result and replies to you in plain English in the conversation.

You never saw Chrome open. You never clicked. You never wrote a line of code. You just asked in plain English, and you got the answer in plain English. That's it.

What the script looks like (just so you know)

You'll never write this yourself, but here's what Claude types in your terminal when you ask it to grab the title of jerwis.fr:

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.goto("https://jerwis.fr");
      const title = await page.evaluate(() => document.querySelector("h1").innerText);
      console.log(title);
    EOF
    

In plain English: "open a browser, go to jerwis.fr, grab the text of the big headline (the h1), print it in the terminal." Done. The browser closes on its own, the title shows up, Claude hands it to you. Five seconds flat.

5 real cases · even if you're not a dev

On paper, dev-browser sounds like a "dev tool for scraping and testing." In real life, it's an assistant that unblocks tasks you'd otherwise do by hand, badly, wasting time. Here are 5 cases where I use it that have nothing to do with code.

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.goto("https://yoursite.com/pricing");
      const text = await page.evaluate(() => document.body.innerText);
      console.log(text.includes("€49") ? "OK €49 visible" : "MISSING");
    EOF
    

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.goto("https://blocked-site.com/article");
      await new Promise(r => setTimeout(r, 3000));
      const article = await page.evaluate(() => document.querySelector("article").innerText);
      console.log(article);
    EOF
    

Case 03 · Capture a design for inspiration

You spot a beautiful site, you want to share it with your designer · screenshot in 5 seconds

You stumble on a site you love (a competitor, a slick landing page, a freelancer's portfolio). You want to send it to your designer or drop it into a meeting. You tell Claude: "take a screenshot of stripe.com at desktop size." It launches dev-browser, captures the image at 1920×1080, shows it in the conversation, and you save it.

Also handy for archiving a page before it changes. A promo, a legal notice, a version of a competitor's page you want to keep. Dev-browser freezes the moment, and you've got the proof.

The command:

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.setViewportSize({ width: 1920, height: 1080 });
      await page.goto("https://stripe.com");
      await page.screenshot({ path: "stripe-home.png", fullPage: true });
      console.log("Screenshot saved: stripe-home.png");
    EOF
    

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.goto("https://yoursite.com");
      await page.fill("input[type=email]", "test@example.com");
      await page.click("button[type=submit]");
      await new Promise(r => setTimeout(r, 2000));
      const message = await page.evaluate(() => document.querySelector(".success-msg")?.innerText);
      console.log(message || "No success message found");
    EOF
    

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.setViewportSize({ width: 390, height: 844 });
      await page.goto("https://yoursite.com/pricing");
      await page.screenshot({ path: "pricing-mobile.png", fullPage: true });
      console.log("Mobile captured");
    EOF
    

My 3 favorite commands · ready to copy

You don't need to memorize these commands. You can copy them, paste them into your terminal swapping out the URL, and you're off. Or better: show them to Claude once and say "this is my command style, keep it," and it'll reuse them on its own from then on.

Command 01 · Screenshot of a whole page

The daily bread. Capture a full site, top to bottom, at default desktop size.

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.goto("https://jerwis.fr");
      await new Promise(r => setTimeout(r, 2000));
      await page.screenshot({ path: "capture.png", fullPage: true });
      console.log("Screenshot saved: capture.png");
    EOF
    

The capture.png file shows up in the folder where you ran the command. You change the URL, run it again, and you get a new capture.

Command 02 · Check that a page contains a word or a number

Great for confirming a change is live without opening a browser. Answer in 1 line in your terminal.

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.goto("https://jerwis.fr/articles");
      const text = await page.evaluate(() => document.body.innerText);
      console.log(text.includes("dev-browser") ? "OK found" : "ABSENT");
    EOF
    

You swap "dev-browser" for any word, price, name. Handy for checking that a price change is rendering, that a new article shows up in the list, that a promo is actually visible.

Command 03 · Submit a form and capture the result

The newsletter, contact, or order signup test. Fill in, submit, check the confirmation message.

dev-browser --headless <<'EOF'
      const page = await browser.getPage("main");
      await page.goto("https://jerwis.fr");
      await page.fill("input[type=email]", "test@example.com");
      await page.click("button[type=submit]");
      await new Promise(r => setTimeout(r, 3000));
      await page.screenshot({ path: "form-result.png", fullPage: true });
      console.log("Form submitted, capture saved");
    EOF
    

You adapt the field selector (input[type=email]) and the button selector (button[type=submit]) to your page. If you don't know them, ask Claude: "look at the HTML of this page and tell me which selectors to use for the newsletter form." It finds them, it adapts the command.

5 bonus uses · that I'm testing right now

Beyond the 5 truly daily cases, here's what I've been experimenting with for a few weeks. Not all of it is polished, but it's already usable if you want to push further.

Install · 3 minutes flat

Prerequisites

• Mac or Linux (Windows Subsystem for Linux works too)
• Node.js already installed (check with node --version · if it prints a number, you're good)
• Claude Code Max (or at least Claude Code free)

Step 1 · Install

Open your terminal. Paste the command Claude gives you when you ask it "install dev-browser for me." It usually goes through npm install -g @anthropic-ai/dev-browser (or whatever the equivalent is for the current version · ask Claude for the exact current command).

Step 2 · Test

Still in your terminal, type:

dev-browser --help
    

If you see a list of options scroll by (something like the screenshot below), it's installed. If not, restart your terminal or ask Claude to help you debug.

Step 3 · First use

Open Claude Code in any folder, and tell it:

"Use dev-browser to go to jerwis.fr and give me the title of the page"

Claude will write a tiny script, run it, and hand you back "Follow AI. Without being a dev. Without spending your evenings on it." (my current H1). If it works, you're set. If not, Claude will help you fix it.

The 3 limits I've run into

Limit 1 · Sites that detect bots

Some sites (X/Twitter, logged-in LinkedIn, Reddit sometimes) detect automated browsers and block them. Dev-browser can connect to your existing Chrome with the --connect flag (a flag = an option you pass to the command), but that's more technical. For the general public, I simply accept that 10% of sites refuse · I move on to something else.

Limit 2 · Very JavaScript-heavy pages

Some pages take 5-10 seconds to fully load. By default, dev-browser waits 2 seconds. If Claude can't find an element, it's often a timing issue. The fix: ask Claude to add await new Promise(r => setTimeout(r, 5000)). It'll know what to do.

Limit 3 · Not the answer if you want to scrape at scale

Dev-browser is built for one-off tasks · checking a page, testing a flow, grabbing one precise piece of data. If you want to scrape 10,000 pages for a massive project, you need other tools (Apify, Bright Data, or custom code — built to order). For everything else, dev-browser is enough.

My personal rule: as long as I stay under 50 pages at once, dev-browser does the job without flinching. Beyond that, I switch to a dedicated tool or split it into several batches. 95% of my uses fit comfortably within that limit.

FAQ dev-browser.

Where to go from here

If you want to dig deeper:

1. Claude Code · the full setup — to install before dev-browser, obviously
2. The Claude Code workflow tips — the logical next step to structure your sessions and combine dev-browser with other habits
3. Claude Code loops, explained — what you trigger on a loop when dev-browser does its thing (monitoring, watch)
4. My Tools page — the 6 tools I use alongside Claude Code every day

If you install dev-browser this week and use it for a real case, reply to my newsletter and tell me about it. It helps me know what works for other people. I read everything, I reply, and I can be wrong about your needs — your feedback sets me straight.