Give your AI agent a computer in the cloud. A managed cloud sandbox — a Linux desktop with a browser open inside — driven by your choice of LLM (Claude, GPT, or Gemini).
pip install computeruse-sdkfrom computeruse import Sandbox # distribution: computeruse-sdk · import: computerusePreview release. The hosted API is not live yet. This package ships the planned SDK shape so you can read it, type-check against it, and prepare your code. Every method raises
PreviewError. Join the waitlist at computeruse.run.
from computeruse import Sandbox
with Sandbox(model="claude-sonnet-4-6") as s:
s.browser.goto("https://example.com/contact")
result = s.browser.run("Fill the contact form with my info.")
s.computer.run("Open the downloaded PDF in /Downloads and summarize to /tmp/out.txt.")One class. Two surfaces. Model is a kwarg.
s.browser |
Web-page tasks: navigate, click, scrape, fill forms. CDP underneath; the agent picks actions. |
s.computer |
Desktop tasks: open files, run a CLI, edit a document, switch between apps. The browser is one window on the same desktop. |
The two surfaces share the same Linux machine. A file downloaded by s.browser lands in ~/Downloads and is visible to s.computer. Cookies, clipboard, and the residential IP are all shared.
Sandbox(model="claude-sonnet-4-6") # Anthropic Claude
Sandbox(model="gpt-5") # OpenAI GPT
Sandbox(model="gemini-3-pro") # Google GeminiBring your own provider API key via env var (ANTHROPIC_API_KEY / OPENAI_API_KEY / GOOGLE_API_KEY). The SDK forwards it to the in-sandbox agent runner — we never see your model traffic.
Currently, calling any Sandbox.* method raises computeruse.PreviewError pointing at the waitlist. The package exists to publish the planned SDK shape so you can prototype against it.
>>> from computeruse import Sandbox
[computeruse] Preview package — hosted API is not live yet.
Sandbox() and every .run() raise PreviewError.
Join the waitlist: https://computeruse.run/
>>> with Sandbox(model="claude-sonnet-4-6") as s: pass
computeruse.PreviewError: computeruse is in private preview — Sandbox(model='claude-sonnet-4-6').__enter__ is not callable yet. ...Silence the import banner with COMPUTERUSE_QUIET=1.
Existing options for running an AI agent against a real computer:
- Browserbase, Browser Use Cloud — give you a browser in the cloud. No desktop. If your agent ever needs to open a file outside the browser, edit a document, or use anything that isn't a web page, you're stuck.
- Anthropic Computer Use reference Docker — a self-hosted single-tenant container. Great for tinkering, painful for production (no concurrency, single VM, you build the IP rotation and CAPTCHA contract).
- E2B Desktop — closest peer. General-purpose code-interpreter sandbox that exposes a desktop. We are the same primitive narrowed and packaged specifically for AI agents driving browser-plus-desktop tasks, with a unified
Sandbox.browser/Sandbox.computerSDK surface. - Self-hosted Playwright + a VM — cheap at compute, expensive in ops time.
Computer Use Cloud is one Sandbox that gives you both a browser and a desktop, drives by your choice of model, with one Python SDK.
| Component | Status |
|---|---|
| Website | Live · computeruse.run |
| Python SDK shape | Published (this package, v0.1.0 preview) |
| Hosted API | Private preview · waitlist open |
| Apache-2.0 runtime | Planned · M5 |
See SPEC.md for the engineering contract — runtime architecture, the per-provider routing, metering, live view, phased roadmap.
- Homepage: https://computeruse.run/
- Waitlist: https://computeruse.run/#signup
- Engineering spec: SPEC.md
- Strategy / IA / roadmap: computeruse-run/specs
- Publish flow (for maintainers): PUBLISHING.md
Apache-2.0. See LICENSE.