Orgo - Cloud computers for AI agents

Orgo exposes programmatic endpoints for every computer action: screenshot, click, type, key, scroll, bash. You bring the model, Orgo provides the computer. Any model with computer use support works. Here is how to wire them up.

Coordinate space

Orgo computers boot at 1280x720x24. Claude and OpenAI read the screen size from the screenshots you send. They return coordinates in that same pixel space. Send an Orgo screenshot unresized and the coordinates are already the computer’s pixels, so you can click with them directly. Gemini is different. It returns a normalized 0-999 grid, which you denormalize against the computer’s real resolution:

x = int(fn_args["x"] / 1000 * 1280)
y = int(fn_args["y"] / 1000 * 720)

Downscale a screenshot before sending it, or denormalize Gemini’s grid against 1024x768 on a 1280x720 computer, and every click lands short of where the model aimed. Nothing errors. The agent looks like it is confidently clicking the wrong thing.

If you create a computer at a non-default resolution, or resize its screen later, use that resolution instead. GET /computers/{id}/screens reports the current width and height.

Anthropic Claude

Claude drives computer use through the Messages API. computer_toolset_20260801 declares every member tool in one entry and needs no beta header.

import anthropic
from orgo import Computer

computer = Computer()
client = anthropic.Anthropic()

messages = [{"role": "user", "content": "Open Chrome and search for AI news"}]

tools = [
    {"type": "computer_toolset_20260801"},
    {"type": "bash_20250124", "name": "bash"},
]

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=8192,
    tools=tools,
    messages=messages,
)

# Claude can return several member calls per turn. Run them in order.
for block in response.content:
    if block.type != "tool_use":
        continue

    if block.name == "bash":
        output = computer.bash(block.input["command"])
    elif getattr(block, "toolset_name", None) == "computer":
        if block.name == "screenshot":
            image = computer.screenshot_base64()
            # Return as a tool_result with a base64 image
        elif block.name == "left_click":
            computer.left_click(*block.input["coordinate"])
        elif block.name == "type":
            computer.type(block.input["text"])
        elif block.name == "key":
            computer.key(block.input["text"])
        elif block.name == "scroll":
            computer.scroll(block.input["scroll_direction"], block.input["scroll_amount"])

computer.destroy()

Models: claude-opus-5, claude-sonnet-5, claude-opus-4-8 Tools: computer_toolset_20260801 + bash_20250124 Beta header: none Older models: Opus 4.7 and Sonnet 4.6 use computer_20251124 with the computer-use-2025-11-24 beta Guide: Claude Computer Use Docs: Anthropic Computer Use

OpenAI GPT

OpenAI’s computer use works through the Responses API with a built-in computer tool. The tool takes no configuration, and each turn returns a batch of actions.

from openai import OpenAI
from orgo import Computer

computer = Computer()
client = OpenAI()

response = client.responses.create(
    model="gpt-5.6",
    tools=[{"type": "computer"}],
    input="Open Chrome and search for AI news",
)

# Process computer_call actions, in order
for item in response.output:
    if item.type != "computer_call":
        continue
    for action in item.actions:
        if action.type == "click":
            computer.left_click(action.x, action.y)
        elif action.type == "double_click":
            computer.double_click(action.x, action.y)
        elif action.type == "type":
            computer.type(action.text)
        elif action.type == "keypress":
            keys = action.keys
            computer.key("+".join(keys).lower() if len(keys) > 1 else keys[0])
        elif action.type == "scroll":
            scroll_y = getattr(action, "scroll_y", 0)
            computer.scroll("down" if scroll_y > 0 else "up", max(1, abs(scroll_y) // 100))

computer.destroy()

Models: gpt-5.6, gpt-5.5 Tool: computer Older integration: computer-use-preview with the computer_use_preview tool Guide: OpenAI Computer Use Docs: OpenAI Computer Use

Google Gemini

Computer use is a built-in tool on Gemini 3.x, called through the Interactions API. Declare environment: "desktop" for an Orgo computer.

from google import genai
from orgo import Computer

computer = Computer()
client = genai.Client(api_key="YOUR_GEMINI_API_KEY")

# Gemini requires PNG; Orgo returns JPEG. The guide below has the
# get_screenshot_png() helper that converts and re-encodes to base64.
screenshot_png = get_screenshot_png(computer)

interaction = client.interactions.create(
    model="gemini-3.7-flash",
    input=[
        {"type": "text", "text": "Open Chrome and search for AI news"},
        {"type": "image", "data": screenshot_png, "mime_type": "image/png"},
    ],
    tools=[{
        "type": "computer_use",
        "environment": "desktop",
        "enable_prompt_injection_detection": True,
    }],
)

# Coordinates are normalized 0-999, so scale them to the screen
for step in interaction.steps:
    if step.type != "function_call":
        continue
    args = step.arguments
    if step.name == "click":
        computer.left_click(int(args["x"] / 1000 * 1280), int(args["y"] / 1000 * 720))
    elif step.name == "type":
        computer.type(args["text"])

computer.destroy()

Model: gemini-3.7-flash Tool: computer_use with environment of browser, desktop, or mobile SDK: google-genai 2.7.0 or later Coordinates: normalized 0-999, scale to pixel resolution Legacy: gemini-2.5-computer-use-preview-10-2025 with generate_content Guide: Gemini Computer Use Docs: Gemini Computer Use

Other models

Any model that outputs structured actions (click, type, screenshot) can work with Orgo. The pattern is always the same:

Create a computer with Computer()
Screenshot with computer.screenshot_base64()
Send the screenshot and instruction to your model
Execute the model’s actions with computer.left_click(), computer.type(), and so on
Loop until the task is done

Open-source computer-use agents follow the same pattern. Agent S3 is one worked example.

Orgo action reference

Action	Python	Description
Screenshot	`computer.screenshot_base64()`	Capture screen as base64 JPEG
Left click	`computer.left_click(x, y)`	Click at coordinates
Right click	`computer.right_click(x, y)`	Right-click
Double click	`computer.double_click(x, y)`	Double-click
Type	`computer.type("text")`	Type text
Key press	`computer.key("Enter")`	Press key or combo (`ctrl+c`)
Scroll	`computer.scroll("down", 3)`	Scroll direction and amount
Bash	`computer.bash("ls -la")`	Run terminal command
Wait	`computer.wait(2)`	Pause in seconds

Full API details: API Reference

Get Started

Guides

Migrations

Templates

Model integrations

Use Any Model

Coordinate space

Anthropic Claude

OpenAI GPT

Google Gemini

Other models

Orgo action reference

​Coordinate space

​Anthropic Claude

​OpenAI GPT

​Google Gemini

​Other models

​Orgo action reference

Coordinate space

Anthropic Claude

OpenAI GPT

Google Gemini

Other models

Orgo action reference