Skip to main content
Back to blog
•2 min read

We Shipped Computer Use

AIanthropicclaude

october 22nd: we announced claude computer use.

claude can control a computer. point, click, type, navigate. like a person.

i watched this feature ship from the inside. the AI future is arriving fast.

what it does

you give claude access to a virtual desktop. it can:

  • move the mouse
  • click on things
  • type text
  • navigate interfaces
  • complete multi-step tasks

show it a web form? it can fill it out. ask it to find information in an app? it can browse for you.

it's agentic AI in practice. not a chatbot. an operator.

the development

i didn't work on this directly (still ramping on alignment science stuff), but i watched the team work on it.

the complexity is immense. teaching a model to understand visual interfaces. to plan actions. to handle unexpected states. every desktop app is different.

the fact that it works at all is impressive. the fact that it works well is remarkable.

the implications

this is a step toward AI that actually does things, not just suggests things.

current AI: "here's how you would do X" agentic AI: "done, i did X"

the efficiency gains are obvious. but so are the risks.

what if it clicks the wrong thing? what if it misunderstands instructions? what if it acts in ways you didn't expect?

these are the questions we think about constantly. safety isn't an afterthought here.

responsible deployment

we released this as a beta. experimental. APIs for developers to build on.

we're being careful. limiting scope. building safeguards. watching for misuse.

this is the anthropic approach: capability matched with caution.

watching from inside

seeing a major feature launch from inside the company is different than watching from outside.

from outside: "cool! claude can use computers!"

from inside: months of work. debates about safety. testing edge cases. last-minute fixes. anxiety about reception.

then it launches and... the world responds. and you wait to see if everything works as intended.

what this means for the field

computer use is one of many agent capabilities that are coming.

web browsing. code execution. tool use. memory. planning.

we're building AI that doesn't just think—it acts.

the future is interesting. and slightly terrifying. in a good way? i think?

watched the announcement livestream from my desk. surreal to hear about features i've seen in internal demos. this is my life now.