Browser sandbox for AI agents

A Real Browser for Your

An isolated browser sandbox your agent can actually click — billed by the second.

$0.10 / hr Pay as you go
No card, no pairing codes Sub-second browser launch Human takeover built-in
amazon.com/dp/B0CL93RNQ8
@e3 Wireless mechanical keyboard product photo

Wireless Mechanical Keyboard – Hot-Swap RGB, 75% Layout

★★★★★ 2,847
$129.99 $189.99
FREE delivery Tuesday
$129.99
In Stock
@e5 Add to Cart
@e7 Buy Now
Lease live · 47 ms
Takeover ready
Real-world automations

From E2E tests to your
company's intranet — same one command

E2E Testing

Replace your flaky Selenium suite

Describe the flow once — agent runs it and screenshots every step.

$instbox browser snapshot && click @e5
Enterprise SSO

Walk past Okta, Duo, and 2FA

Hand the SSO challenge to a human for 10 seconds — then resume.

$instbox browser takeover
AI commerce

Let agents actually buy things

Real Chrome, real cookies — Stripe and captchas see a normal shopper.

$instbox browser click @addToCart
Form automation

Submit 500 forms before lunch

CSV in, screenshots out — one sandbox per row, all in parallel.

$browser fill @claim_id INS-48291-A
Data extraction

Slip past Cloudflare and DataDome

Real Chrome fingerprint — bot-walls wave it through every time.

$instbox browser snapshot --json
Session resume

Pick up exactly where you left off

Disconnect now, resume tomorrow — --resume restores every tab, cookie, and login from the encrypted snapshot.

$instbox browser launch --resume
Slot grabbing

Grab the visa slot the second it opens

Loop snapshot every second — agent fires click the moment a slot drops.

$instbox browser snapshot && click @slot
Price tracking

Compare 20 sites — buy when it dips

20 sandboxes in parallel, JSON snapshots back to your agent every minute.

$instbox browser navigate && snapshot --json
Change detection

Wake me when this page changes

Hourly screenshot — your agent diffs and pings Slack on a delta.

$instbox browser screenshot --output now.png
How it works

From prompt to page —
in four CLI calls

Your Agent

Claude, GPT, OpenClaw — anything.

Any LLM

browser launch

One call · isolated Chrome leased.

<1s start

Drive the Browser

snapshot · click · fill · eval · takeover.

@ref + Shadow DOM

Results & State

JSON back to your agent · state saved for --resume.

JSON everywhere
See it in action

Drive the browser, or hand it over

sunsetstays.com/deals
SunsetStays DestinationsDealsSign in

The best travel for less

Snapshot → click @e5 → fill checkout. When the booking site throws a CAPTCHA, your agent calls takeover and the human takes 10 seconds.

Book your trip
lease_a7c3f9… Browser live
Tool agent-browser v2.4
Action latency 47ms
Snapshot 142@refs
Lease cost $0.00 in free 30m
Takeover READY
[14:02:18] browser launch → lease_a7c3f9
[14:02:19] browser navigate sunsetstays.com
[14:02:19] browser snapshot → 142 @refs
[14:02:20] browser click @e5 → ok
[14:02:21] browser screenshot → 412KB

Install in One Command

Pick the method that fits your setup. All options create a symlink in ~/.local/bin/instbox automatically.

  1. 1

    Sign in

    Bake a one-time token into your command so the CLI works on first run. Sign-in opens in a small popup — the token never leaves this browser session.

  2. 2

    Where should we install?

  3. 3

    Run this in your terminal

    Default
    
                  

instbox.zip

Top-level instbox/ directory · SKILL.md + bin/ + lib/

Anonymous Download
OpenClaw prompt

        

Pick the location you installed to. We'll show the matching uninstall command.

Put your AI agent in a real browser.

One curl | bash, an OIDC sign-in, and your agent has a browser it can actually click.