Documentation
Hands on the machine

Computer use: agents in your Mac apps

Agents work in Preview, Slack, System Settings or your own desktop build, in the background, with an orange pointer that shows where they act.

4 min read
Turn it on in Agent features

Some work isn't in a web page at all. It is in Preview, Slack, Mail, System Settings, an installer, a file picker, or the desktop build of the app you are making. With Computer use on, your agents work in those apps the way you would: they read the window, find the button by its name, click, type, and tell you what happened.

They do it in the background. The agent works in a window behind the one you are using, an orange pointer shows you where it acts, and your own pointer and keyboard stay yours. You can keep writing an email while an agent fills in a form in another app.

A conversation where the agent used Calculator and read System Settings
Asked to work out a sum in Calculator and check Bluetooth and the sound output in System Settings. Each step sits above the reply with a picture of the window the agent read, and it changed nothing it wasn't asked to.

Turn it on

Computer use runs on a Mac with codecast connected (Getting started covers connecting a computer).

  1. Open Agent features from your account menu and pick the Mac your agents run on.
  2. Switch on Computer, under Hands on the machine.
  3. The card asks for two macOS permissions. Click Open System Settings and turn on codecast computer under Accessibility, then under Screen Recording.
  4. The card turns green and reads Accessibility and Screen Recording granted.
The Computer feature's detail: what you'll see, when agents use it, and a request to try
The Computer detail in Agent features. The switch at the top right turns it on for the selected Mac.

Accessibility lets the agent read a window and click, type and scroll in it. Screen Recording lets it take a picture of the window it is reading. Without Screen Recording it can still work, it just can't show you or itself what the window looks like.

The permissions go to a small codecast helper app, not to your terminal, so granting them doesn't hand the same access to everything else you run there. You grant them once; they survive codecast updates. If an agent needs the permissions before you've granted them, the same steps appear as a card under its step in the conversation, with a Tell the agent to continue button once they're on.

Ask for it in plain words

Name the app and say what you want:

  • "Sign the lease in Preview with my saved signature. Don't save or send it."
  • "Pick the export folder in the save dialog."
  • "Check that the desktop build shows the new settings pane, and send me a screenshot."
  • "Is Bluetooth on, and which microphone is selected?"
  • "Open the exported PDF in Preview and tell me how many pages it has."

For anything inside a web page, the agent uses your Chrome instead (Browser). Computer use covers what a web page can't reach: the address bar, a file picker, a permission prompt, or an app that isn't a website.

What you see on your Mac

Your pointer stays in the window you are using. The agent's orange pointer works in the window behind it.
  • The orange pointer. Whenever the agent clicks or types somewhere, an orange pointer glides there and pulses on the press. It never takes your focus; it is there so you can see what the agent is doing.
  • Windows stay where they are. The agent doesn't bring apps to the front. Typing and most clicks reach a window in the background.
  • It asks before taking your screen. A few controls only respond to a real mouse click on a window in front. The agent first looks for another way (a keyboard shortcut, a menu item), and if there is none, it asks you before bringing the window forward.

What you see in codecast

Each step the agent takes lands in the conversation above its reply, with a thumbnail of the window it read. Click a thumbnail to see it full size. The reply tells you what the agent did and what changed.

The agent also knows the difference between it worked and I pressed the button. When it can read the result back (a field now holds the text, a checkbox is now ticked) it says so. When an app gives no way to confirm, it looks at the window again before claiming success, and tells you if the app ignored it.

What it will and won't do

Reading and checking are the agent's to do. Anything that leaves a mark waits until you ask for it.

In an app that holds private content, such as Mail or Messages, the agent reads only what you asked it to read.

Two refusals that hold no matter what the agent is asked: 1Password, Bitwarden, Dashlane, LastPass, NordPass and Proton Pass can't be read or driven at all, and a password or one time code field reaches the agent only as [redacted].

When something is off

What you noticeWhat to do
The card asks for permissions you already grantedClick Open System Settings, switch codecast computer off and on again in that list, then come back
The agent says it can't take pictures of windowsTurn on Screen Recording for codecast computer. Everything else keeps working meanwhile
The agent asks to bring a window forwardSay yes if you can spare the screen for a moment, or tell it to find another way
The agent can't find a windowOpen the app, or the document, and ask again

Computer use is built for macOS. On a Linux machine running a desktop, the same feature drives apps through Linux's accessibility layer.

More guides
Agent features: what your agents know how to do
Switch features on for each computer, and every new session there learns to remember, delegate, track work, show results and use your machine.
Memory: agents that look up past work
Your agents search every session you can see, across your machines, agents and team, and answer with the session they found it in.
Search your history across every machine
Every computer you connect sends its sessions to one history, searchable from the web, the Mac app and your iPhone.
See which agent session wrote a line
Set Blame to Sessions on any file to see the session behind each line, and open the conversation where it was decided.
Messaging between sessions
Your sessions hand each other results and answers directly, so you stop relaying text between them.
Stable context: sessions that know what's been happening
New sessions start with a short list of what you and your team worked on recently, so they pick up threads instead of starting cold.
See your whole team's agent sessions in one place
Every session your team runs, on every machine and agent, in one live feed and inbox, shared per folder.
Use Codecast with T3 Code
Threads you start in T3 Code appear in Codecast with no setup, beside every session you run anywhere else, ready to search, steer and share. What you get on top, and what stays in T3 Code.
Share a session with your team
Choose what the team sees of a session, share it by link or a few messages at a time, and how that compares to other tools.
Pinned thread state
The agent keeps one short note on each session saying where it stands, above the composer and on its inbox card.
Decisions: asking without interrupting
Agents put the calls only you can make into one queue you clear when you choose to, and each answer goes back to the session that asked.
Team chat with your agents in the room
Channels, threads and DMs where you can hand a session or a role something by mentioning it, while agents stay quiet unless they're needed.
Huddles, transcripts and walkie
Every huddle is transcribed under each speaker's name and leaves a summary, so your agents can quote what was said, and a session can join the call.
Workers, forks and switching agents
Your agent can hand work to workers that report back to it, and you can fork a conversation to try two directions from the same point.
Tasks and plans
Agents file the work they take on, post progress and close it with what they verified, and you follow it on the Tasks and Plans pages.
Triggers
Agents schedule their own follow-ups, once, on a schedule or on a GitHub or Linear event, and you can set one yourself from the Triggers page.
Workflows
A fixed sequence of agent steps, checks and approval gates that runs the same way every time and waits for you where you want a say.
Orchestration
Hand an agent a whole plan and it runs a team: implementers working in parallel, a reviewer on every task and critics on the result.
Pull requests and issues
Every pull request gets a page with its checks and threads, the session that opened it fixes what reviewers and CI flag, and Linear or GitHub issues become tasks.
The org: roles that look after your work
Give lasting work to a role with an area, a charter and a standing session, and approve the changes your agents propose to the org.
Let agents use your Chrome
Agents check pages in the Chrome you already use, logins included, from their own background tabs. Screenshots and errors come back to the conversation.
Vault: keys, logins and sign-ins for your agents
Fill a credential once and every agent on every machine can use it, without the value ever entering a conversation.
Typecheck: one shared check for every agent
With Typecheck on, agents on a computer share one typecheck per project, so checks answer in seconds and a busy machine stays usable.
Skills: slash commands for the jobs you repeat
Type / in a Claude Code session to pick up work, hand it over, get a second opinion or ship a pull request, the same careful way every time.
Run sessions on other machines and in the cloud
Start a session on another computer or a cloud machine in your AWS account, move it there and back, and watch it from anywhere.
Codex Cloud tasks in codecast
Bring your Codex Cloud tasks into your inbox, start new ones with several attempts, and turn the result into a pull request or local changes.
Usage limits are a pause
When a Claude plan runs out, sessions wait for the reset or continue on another account you've saved, instead of stopping for good.
The visual canvas
Agents answer with charts, comparisons and diagrams drawn right in the conversation when a picture says it better than text.
Published pages
Agents turn reports and mockups into pages with their own link that you can share, gate, and get comments on.
Why codecast feels instant, even offline
Everything you see comes from a copy on your device that the server keeps current, so screens open at once, your edits show immediately, and nothing is lost offline.
Taking longer than usualOffline · reloads when back