Introducing: Browser Use CLI 2.0 🔥 The most efficient browser automation CLI tool > 2x the speed, half the cost > Easily connect to running Chrome > Uses direct CDP Try it now 🔗↓

Try it here:

direct CDP is the right call been burned by playwright overhead way too many times does the 2x speed hold when youre running multiple concurrent sessions?

Yes!

The Claude Chrome extension is free to use with Claude CLI. Honest question: what does this improve over that setup?

1) the Browser Use CLI is also free 2) Fast and token efficient 3) Don’t need to install a Chrome extension 4) Can be used with any CLI agent

@DanielMuvdiYT What is your pricing page for then?

@DanielMuvdiYT Our cloud platform

the question everyone is wondering: can it run headless and do the same thing as this video demo?

Of course!

Benchmark: browse vs agent-browser vs browser-use It's definitely a bit better than agent-browser, but would't say it's the "fastest" or the most token efficient. benchmark repo: browse: browser-use : agent-browser: Disclaimer: I've built browse

My problem is not how to use this. My problem is now what to choose between Playwright MCP, WebMCP, Puppeteer MCP, Browser-use MCP, Chrome DevTools Protocol MCP, and agent-browser. It's becoming so confusing.

More efficient than agent-browser?

Here's Claude Code + Browser Use CLI filling out a form 😉

@davidtsolheim How can Browser Use CLI be integrated into Claude Code? skills? mcp?

@davidtsolheim Download the CLI, then download the skill to your CLI coding agent Here are the docs

half the cost? isn’t this running locally?

Half the tokens! When using Browser Use CLI with Claude Code, we are 2x more token-efficient than others!

gotcha! not for token-maxxers then. 😂

OpenCLI will cooperate with it well it turn all website to be CLI

"2x the speed, half the cost" compared to what lol

compared to other browser automation CLIs!

I love you guys but you need to be more specific. The variability in speed and cost among browser automation CLIs is huge, so "2x the speed, half the cost" seems like an arbitrary comparison again a broad set of tools

Here's what makes CLI 2.0 actually different — in plain terms: Old browser automation tools (Playwright, Puppeteer) work by launching a brand new Chrome every single time. Fresh tab, no cookies, no login sessions, no memory. Great for testing. Terrible for AI agents doing real work. Browser Use CLI 2.0 connects directly to the Chrome you already have running — via CDP (Chrome's native "remote control" protocol). That means: → Your agent can see your Gmail, Notion, dashboards — anything you're already logged into → No re-authentication overhead on every run → 2x faster because it skips the browser launch entirely → Half the cost because fewer tokens wasted on setup context The shift happening here: browser automation is moving from isolated sandboxes → live session control. Playwright was built for developers testing apps in clean environments. Browser Use CLI 2.0 is built for AI agents working inside your actual digital life. Two very different tools solving two very different problems. The second one is where agentic AI actually gets useful.

can it upload documents in a form?

Yes! The CLI excels on our interactions dataset, especially form filling

Cool! i will definetly try it out then!

Direct CDP is powerful, but maintaining those selectors is the real bottleneck.

‘browser-use state’ is super optimized for agents!

this is like people initially shilling openclaw to clean their email 😭 could have been a bit more creative Like help me commit tax fraud agentic

Direct CDP is a nice call. The “connect to running Chrome” bit usually makes automation feel less brittle than spinning fresh sessions every run. I keep going back and forth on CLIs vs SDKs though.

2x speed half the cost and it reads my notifs too, what am i even needed for at this point

Lightpanda Support?

explains all the AI replies I've been getting lately

Browser Use 2.0 looks insanely fast. love the direct CDP speed. One thing I keep seeing is agents still guessing how sites actually behave: async timings, irreversible actions, safety rules, error states. That’s exactly why I shipped operate.txt. It’s a static YAML at /operate.txt (robots.txt style) that gives agents the behavioral contract. What flows break your agents most often? Throw ideas here and I’ll add real examples to v0.3 + credit contributors.

That's a banger of a post on your feed.

Me to my manual browser after seeing Browser Use 2.0

那本地都能做了，它家的 cloud 服务可以做什么？你们谁试过

Cloud gives you access to our agent-ready browsers and browser subagents! Massive parallelization

Is there a way to get rid of the manual confirmation in the browser? It ruins the whole afk experience

For now, you need to click it at least once Our goal is to make AI browser automation completely frictionless!

I’m pretty sure it resets tho after a while and you have to click again in the next session, right? Btw is there a solution you offer that can I can have remote browser logged in to all my accounts with static ip? I have a sub and paid for credits but I can’t find a usecase yet

@browser_use Yes, you can buy your own static ip proxy (I recommend brightdata) and use it as a custom proxy:

@browser_use Fantastic! If I can use a custom proxy I’ll just host my own, that’ll give it my real ip 🤌

Faster, cheaper browser automation isn’t just incremental—it’s a step toward agents that can operate the web as efficiently as humans, at scale.

Direct CDP instead of driver abstractions was inevitable once Chrome stabilized enough. The 2x speed claim is interesting — does it hold on JS-heavy SPAs where you still need waitForNetworkIdle-type guarantees? That's always where the high-level wrappers earned their keep.

Is this opensource?

Yes! Here’s the repo:

Thank you 🙏

Thank you for this, this is fireee!

been using browser use for web scraping. 2x speed is huge for agent workflows that need to hit multiple pages per task. the cost reduction makes it viable to run browser agents in production loops instead of just one-off demos.

Should one disable the playwright mcp if using this because coding agents quickly go back to using playwright.

try out the CLI and compare!

Do you have session management / tab management support?

@llmgram

fantastic, but i am curious about the principle. Does the claude do the image detection? or the Web element content comprehension. I mean the latter will be efficient and powerful

The coding agent that uses the CLI handles processing the content, the CLI handles context and executions Each command shares ~the same implementation as our battle-tested tools in the browser-use library

Cool. And recently I read about Webmcp, which is related to build standard web context protocol for AI comprehension. I try and I found using the this types of paradigm can make agent execution more accurate. Does the CLI consider to adopt this?

Why not just Agent-Browser by Vercel it's free and open source? Does this do something better?

this is also free and open-source and arguably better in some ways 😉

In the controlled chrome I’ve tested it was so much slower then agent-browser.