Skip to main content
OpenAI’s computer tool lets GPT see a screenshot and reply with clicks, keystrokes, and scrolls. This guide runs it on an Orgo computer, so the agent gets a full cloud computer and nothing runs on your machine.
The video walks through the starter repo from OpenAI’s docs to a working agent, first on a local Docker computer, then on Orgo.

Pick how to run it

Let Orgo run the loop

The Computer Use API runs the whole agent on Orgo’s side. Send the task and a computer_id. The agent screenshots, clicks, and types until the task is done, then returns its final message in OpenAI’s chat completions format. Replace $COMPUTER_ID with the UUID of a running computer.
orgo-agent runs GPT-6.1 Terra on its fast tier, billed at twice Terra’s standard rates. Set "stream": true to receive progress as Server-Sent Events.

Run the agent loop yourself

Set up

The complete example

This script creates a computer, runs the loop, and deletes the computer when it finishes. It calls the Orgo API over HTTP, so it needs no Orgo SDK. The script creates the computer on your plan. A free account has no plan computers, so add "billing_mode": "hourly" to the create body. Orgo then launches a Medium computer billed by the hour from your credit, as described in Create computer.
agent.py
Run it:
Open the printed link to watch the agent work. To keep the computer afterwards, remove the DELETE call in finally.

How the loop works

  1. Send the task with tools=[{"type": "computer"}].
  2. The response holds a computer_call with a batch of actions.
  3. Run every action in the batch, in order.
  4. Send back one screenshot as a computer_call_output, chained with previous_response_id.
  5. Repeat until a response has no computer_call. Its output_text is the answer.

How each action maps to Orgo

Coordinate space

A computer created without a template, like the one in the example, boots at 1280x720x24. A macos computer boots at 1920x1080x24, and a template launch keeps the template’s own resolution. The computer tool takes no display size. The model reads the size from the screenshots you send and returns x and y in the same pixels. Send each screenshot unresized with "detail": "original", and the coordinates are already the computer’s pixels.
If you downscale a screenshot and click the raw coordinates, every click lands short of where the model aimed. Nothing errors. If you must resize, keep the scale factor and map each coordinate back to the original resolution.
The base64 screenshot response includes width and height, and GET /computers/{id}/screens reports the current size of each screen if you created the computer at another resolution or resized its screen later.

Go faster with fast mode

Add service_tier="fast" to every responses.create call. OpenAI’s fast tier runs up to 2.5x faster and bills at twice the standard rate.

Migrating from the preview tool

OpenAI shut down the computer-use-preview model on July 23, 2026, as listed on its deprecations page. An integration still built on it has to move to the computer tool. The biggest code change is the action handler: loop over actions instead of reading one action, and send one screenshot after the whole batch. Delete any hard-coded display_width and display_height, because the model now reads the size from the screenshots.

Safety

The model acts on whatever the screen shows, so treat page content as untrusted input. Keep a human in the loop for consequential steps, as the example does for pending_safety_checks. An Orgo computer handles isolation: anything the agent does stays on that one computer, and deleting the computer ends it.

Next steps

Starter repo

The code from the video, with a local Docker computer and Orgo

OpenAI computer use docs

The computer tool, actions, and safety guidance

Computer Use API

Let Orgo run the agent loop for you

Claude Computer Use

The same setup with Claude