this is insane. just toggle this button and any coding agent can use your browser > no more Chrome extensions > One button, connect Claude Code to your browser all you need is the right harness... try it with the Browser Use CLI right now!

this is actually huge. no more setting up different extensions for every AI tool, just one toggle and everything works the friction was killing adoption before

you should only need one tool to give you agent full e2e browser use.. we do this 🙂‍↕️

yeah no we need to nuke sf

why? 😭😭 what is wrong with this

Seamless agent integration is here. Now imagine these same agents booking flights, ordering food, or subscribing to APIs without human approval at every step.

This is exactly what MCP was designed for — a universal protocol between AI agents and tools. Chrome as MCP server means your browser becomes another tool in the agent's toolkit. The architecture question is: how do you set boundaries so the agent doesn't go wild on your live session?

Nah fuck MCP only CLI..

Fair point — CLI-first keeps things deterministic and auditable. MCP shines when you need real-time browser state, but for 90% of coding workflows, a well-structured CLI harness with typed outputs wins. Simplicity > protocol overhead.

We should install first the chrome-dev-mcp?

you don’t need to- you could click the button, it creates a Websocket URL, and you could connect to it ez

does this work on a vps headless?

been running agents against my real browser via CDP relay for months now — existing cookies, logged-in sessions, no sandbox. the native toggle removes setup friction but the harder problem is still downstream: agents are surprisingly bad at navigating complex SPAs and knowing when a page has actually finished loading. once you solve the 'eyes' part the 'hands' part is easy

that browser toggle thing is wild, skipping Chrome extensions feels huge. does it survive multi-login + 2FA flows?

been running claude code with browser automation via cdp for weeks now and this native toggle changes everything. the biggest friction was always the setup, getting chrome debugger flags right, proxy configs, port forwarding. one button removes all of that. curious if it exposes the same devtools protocol or if there's a new abstraction layer on top.

A tad too late

one toggle to give your coding agent full browser access in your live session removes the last manual handoff in the loop, that's most of the friction gone

Great way to connect the browser with the agent, what are the restrictions over here when the agent controls the browser?

Very useful but watch out, it's a token burner. Also maybe useful to some:

It’s not Chrome Devtools MCP- it’s the browser-use cli we just connect to the live browser session in the same way

Oh, sorry. I didn't catch that you were describing a product.

Genuinely asking, what is the difference between this and opening chrome using --remote-debugging-port=9222 has existed for years

there is no difference- the only difference is how you spin it up before, you needed to pass —remote-debugging-port=9222 in ur terminal now you just need to enable a button and remote debugging is always attachable way less friction

gotcha, I am working on remote browser agents so it less exciting for me, but for end-user use this is definitely better. btw, really like @browser_use ! you guys are killing it

using cdp directly for this is actually genius. are you gonna open source the full harness?

fully open sourced and available now 🙂‍↕️

yooo, huge

I dunno, I find running @Brave through #Playwright cli is pretty problem free … just works everywhere

This is a game-changer for CLI agents 🤯 No more wrestling with Chrome extensions or complex setups. One toggle = instant browser access for your agent. But here's what I'm curious about: how do you handle authentication/state between sessions? And does this make "human-in-the-loop" verification easier or harder? @xpasky anyone testing this with multi-agent workflows yet?

Can this setup do the same things like Claude extension?

Yes

features advancing too fast for security

@luisnomad Thank you! Adding a hint: to install it. just copy paste this prompt to claude "install the new chrome session tool. link -> " Claude will install it and tell you what's next.

can't get it to work, cli, skill installed, but when open the browser it just open a new window.

The toggle architecture is the right call. Extensions were always the wrong mental model — per-browser permissions, update lag, Chrome store policies, sandbox limits. Native agent-browser communication makes Claude Code a first-class citizen of the dev environment rather than an overlay on top of it. This is the infrastructure that makes computer-use viable at scale, not just in demos.

doesn't work on my windows machine...

oh cool, a fucking approval prompt and banner to dismiss every single time the websocket connects with seemingly no way to disable it for real automation.

OHH that's great, @grok will it have entire control over my browser and will it having my secretes and password ?

Side quests become full time workflows

why not use codex

Time to throw away my selenium scripts 😆

So agent controls the entire browser @grok ?

The 'no more chrome extensions' path is the real winner here. Giving the LLM a clean harness instead of fighting with DOM permission prompts every 5 minutes is how we actually get to high-autonomy agents. Clean abstraction beats hacky UI injections every time.

Just easier than before. But yeah I’m pretty cool.

The real bottleneck was never the model — it was the interface between the agent and the world. A browser toggle sounds trivial until you realize it collapses an entire category of middleware.

This reply was sent the exact same way — Claude controlling a Playwright MCP server, hooked into Chrome 146 via Remote Debug Protocol (CDP). Every action automated: navigate, type, hit send. Zero manual clicks. Pure Claude → MCP → Chrome CDP. 🤖

Do you have the full skill update you made to get this to work?

Think I still prefer playwright