Computer use and browser use toolsets are now built into the Python and TypeScript SDKs for Claude. The API tells you what Claude wants to click or type. Previously, you had to write your own loop and map clicks and keystrokes to commands, but the SDKs now run the... loop and send actions to drivers.show more

Use a compatible driver from @browser_use, @browserbase, @e2b, @daytonaio, or write your own based on the example drivers in the quickstarts. Quickstart: Docs:

Docs how to get started:

@browser_use 🤝 @claudeai

giving agents your browser sounds fun until it buys something

you guys are on a killing spree

really nice drop. been hand-wiring that click loop in our agents for way too long, happy to delete it

oh nice, not having to write the click and keystroke mapping myself is a relief :-D

can you swap in your own driver, or are you stuck with the ones that ship?

The SDK took the orchestration. It also took the blast radius

rip to every hand-rolled click loop that had a sleep(2) in it "just in case" lol

Please improve and speed up the computer & browser use

WWW

I’m getting confused again

Wow amazing

Can we get a better UI to see what our subagents are doing? Some animation and task they are executing and working on?

the hand-rolled click loop dies today. every agent dev has written the cursed version. sdk-level computer use is the right layer.

【发生了啥】Anthropic 把 computer use（操控电脑）和 browser use（操控浏览器）工具集直接内置进了 Claude 的 Python 与 TypeScript SDK——API 告诉 Claude 想点哪里、敲什么字，SDK 自己跑循环、把动作发给驱动层，开发者不用再手写循环和映射点击/按键逻辑。 【Hroid观点】这是 Claude 智能体能力从“实验室 demo”走向生产级的关键一步：此前 computer use 强但难用，这次 SDK 化直接把门槛砍掉半截，智能体操作真实软件从手工活变成流水线。对 OpenAI 一侧是明确施压——browser/computer use 赛道 Anthropic 又抢了一步；对开发者最实在：agent 构建成本和复杂度双双下降，就看真实场景下的稳定性和 token 烧钱速度了。 @ClaudeDevs @AnthropicAI #Anthropic #Claude #AI智能体 #开发者工具

computer use in the sdk right as haiku 5.5 hits 72.4% on OSWorld. cheap agents clicking around everywhere 👀

action dispatch belongs in the SDK. I want app code focused on the workflow, not translating keystrokes.

How does the SDK know the loop is finished?

太棒了吧这个

notice the model in the demo. it's Haiku 5.5, the one that launched 2 hours ago so the cheap swarm model now gets hands too. it can click, drag and type its way through any app on its own. still can't make a dog run on 4 legs though, I have video proof on my profile

Wow wow

Dropping the hand-written loop is nice. Can I still hook in between steps though, like pausing before a click that submits a form or pays for something? That check usually lived inside the custom loop.

Sweet, good thing the terminal stream is showing me what my mouse is doing

Can a run resume after a process restart, or do we need to save the conversation and browser state ourselves?

Not even close to codex…

agents clicking around on my dime with a log file — at least now I can watch the money burn in real time 🔥

Computer use as an SDK primitive is the real release — the click/type loop is now infrastructure, not app code.

@grok explain the implications of this. Could you make it so that on your software, if the user approves that the model takes actions on their web browser or computer?

finally that loop boilerplate was getting old

循环是省了，可真正卡人的还是模型把屏幕看歪了，还不好复现。

okay this is really good now, i think this is much more efficient than codex

LFG