Use case

A real browser, on a leash.

Chrome and Playwright run inside the same isolated sandbox as everything else — the agent gets a real rendering engine, and your egress policy decides which sites it can actually reach.

01

QA & regression agents

An agent drives a real Chrome instance through your app's critical paths and reports what broke — no headless-in-CI flakiness from a shared runner.

02

Structured web research

Fetch, render, and extract from pages that require real JS execution, not just an HTTP GET.

03

Form-filling & workflow automation

Agents that complete multi-step flows on third-party sites — scoped to exactly the hosts your egress policy allows.

04

Computer-use style agents

Give a model a browser it can see and click, contained in a sandbox instead of your own desktop.

In practice

Launch, drive, screenshot, close.

sandbox = client.sandboxes.create(template="python") browser = sandbox.browser.launch(headless=True) page = browser.new_page() page.goto("https://app.example.com/pricing") page.click("text=Start trial") page.wait_for_selector("#signup-form") result = page.screenshot(path="/workspace/signup.png") browser.close()

Why isolation matters here specifically

A browser is the widest tool you can hand an agent.

A browser can authenticate as you, submit forms, and reach almost anything on the open web — broader access than a shell alone. Default-deny egress means that access is still bounded by an explicit allowlist, not the agent's judgment about which sites are safe to visit.

Give an agent a browser it can't wander off with.

Chrome + Playwright, preinstalled, same egress policy as every other sandbox.

Create an API key See the isolation model

$pip install ovrin