Introducing the new dev-browser cli. The fastest way for an agent to use a browser is to let it write code. Just `npm i -g dev-browser` and tell your agent to "use dev-browser"

there are like 100s browser agent clis (even Garry Tan has one) why use this one? dev-browser is incredibly bitter-lesson pilled. Rather than inventing a new syntax for browser automation, just use the one they already know: Playwright.

the original dev-browser was just a markdown file and a few scripts that called playwright directly. simple, but slow startup and you had to let your agent run arbitrary node scripts. since its release there has been a lot of innovation in the browser skill space (from things like agent-browser) the new cli addresses the two biggest points of feedback from the original: a proper daemon for fast startup and all playwright scripts run inside a sandboxed quickjs vm. your agent writes real playwright code but can't `fs.rm()` because it had a bad day.

fittingly we just hit 4k stars on github! Check it out today!

sawyer you're a legend bro i think you forgot to mention something absolutely critical here. this isn't using a playwright dev environemnt that doesn't store your sessions, passwords, etc reliably. this is the real browser. and its fast. that alone makes it special.

did you try playwriter?

i did not! i had not realized that you went full cli as well!

cli is much better. MCP has always some issue. you just need a dependency using a console.log and it breaks everything

that is FAST sure you get this all the time but - how do you compare this vs agent-browser from vercel?

here is a video of agent-browser doing the same task for comparison

hahahaha I take your point

hell yeah buddy

The legend is back

i didn’t know i was gone lol

REALLY awesome work brother! Can I send a donation to the github? I LOVE this man.

@kyuxloll no need my dude. just let me know any issues you run into

Definitely bro! I would like to donate regardless, can raise $1,000+ for you man. Just let me know :)

@alanaagoyal About a week ago this was released - would love your comments on this as well @sawyerhood

any plan to have it for poor windows slaves:)

yes! i need to boot up my windows machine

haahhaa. Nice! Best of luck.

update i have just released a new version that should support windows!

Can you please make this into a chrome extension so I can just go into my browser and have it do complex annoying things for me

been using CDP websockets to automate chrome for a week straight now. uploaded 3 tiktok videos, sent 97 emails, and replied to this tweet. all from the same node script. we are truly cooked.

The real bottleneck isn't browser speed, it's agents hallucinating what's actually on the page. Does dev-browser give the agent a reliable DOM snapshot, or is it still guessing?

This looks great for standard browser flows. Curious how dev-browser holds up with more complex form-filling though. Does the agent end up needing way more snapshots in those scenarios?

how is this different to agent-browser from vercel and chrome devtools mcp?

the "write code to control the browser" approach is so much better than screenshot-based stuff. way more reliable, less token waste. this is how agents should work tbh

Unfortunately its useless, asked it to go through a spreadsheet of Zillow properties and make a sheet showing price, amenities, walking distance to X, etc, couldn't do anything because of Google/Zillow CAPTCHA.

the problem though is that you have to click "allow debugging" every single time the agent has to run. which makes automations and scheduled jobs impossible.

How does this compare to Browser Use 2.0?

browser agents peaked when we stopped clicking and started coding.

This is excellent, thank you so much! It works like a charm on Windows. Great work!

@threepointone This is sweet

Is there a way dev-browser can use my existing google chrome profile.. so that it can use my login cookies

nice, basically building on the code mode concept, I wonder what it would take to have a user story written in plain english, and have the model spit out playwright that would drive that user story might get closer to solving the whole fragile e2e testing concept that is always such a pain and breaks so easily

Bruhh… thats stupid fast… i like it 🫵

Just tried, this actually works great! So fast and smooth!

How does it compare to browser-use?

been following dev-browser since you launched it — using Chrome MCP for browser automation right now but the 'let the agent write code' approach feels way more natural. planning to switch over

Different from agent-browser of vercel?

cool idea, but security model for arbitrary browsing agents really matters here

fantastic. superfast. rollout. thanks.

Just installed. Works great. Thanks for putting it out there. Another superpower unlocked for OC 👍

The Playwright choice is key. Every other browser agent tool invents its own DSL that the model has to learn from scratch. Playwright already has massive representation in training data — models write it fluently without extra prompting. Real bottleneck was always startup latency + context window waste, not the automation API itself. This solves both.

Best browser use agent

How does it compare with agent-browser?

Great work Sawyer!