Infrastructure for AI agents

Give your AI agents a real sandbox.

Every agent, task or branch gets its own Postgres, Redis or browser to work in — a disposable sandbox on your laptop, a VM or a Kubernetes cluster, all driven from one CLI. Awake on connect, asleep at 0 B.

Get started See how it works

self-hostedno SDKdocker · kuberneteszero depsMIT

Claude agent · sbx

You

Refactor the auth middleware, then make sure the whole suite still passes.

I'll work inside the sandbox — it has the Go toolchain and the repo mounted at /work, isolated from your machine.

sbx · pr-4711

sbx create pr-4711 --spec sandbox.json  # golang:1.26, repo at /work
sbx exec pr-4711 dev go build ./...
sbx exec pr-4711 dev go test ./...  # ← wakes the box

✓ ok · 91 tests, 0 failures · nothing on the host touched

Built and tested inside the box — my code runs on /work, fully isolated, and sleeps to 0 B when I step away.

Snapshot & fork

Seed a database once, then fork it a hundred times — the migration runs a single time.

sbx snapshot · sbx fork

Checkpoint & resume

Park a box mid-run — memory and processes — and bring it back exactly where it left off.

sbx checkpoint · sbx resume

Locked-down egress

Let the agent reach only the APIs you allow — enforced by a proxy, with no route around it.

egress_allow: ["api.openai.com"]

Stronger isolation

Run untrusted code in a userspace kernel or a microVM — not just a shared-kernel container.

--isolation gvisor · kata

Wakes on a socket

Any TCP connection wakes it; a few idle minutes sleep it back to 0 B. No SDK, no wrapper.

psql · go test · curl

One-shot fixture

A sandbox that lives exactly as long as a test — always torn down, even when it fails.

sbx with test-db -- go test

how it works

There is no start. There is no stop.

Opening a socket is the only signal — so an agent’s tools (psql, a driver, a test runner) wake its sandbox without an SDK, and a few idle minutes sleep it back to zero.

Create it

One command realises the spec — a real Postgres, Redis or browser. --template works with nothing on disk.

Forget it

Ignore it and it sleeps to 0 B — a stopped container, volume intact. You never turn it off.

Connect

psql, a pool, Playwright or a test runner opens a socket and it's back in a second. You never turn it on.

built for branches · agents · CI

Deploy anywhere. Manage from one CLI.

Local agents or remote, it runs anywhere docker or kubernetes does — and sbx ui --connect drives every sandbox, wherever it lives.

localGCPAWSAzureself-hostedbare metal

Wakes on a socket

Any TCP connection. No SDK, no client library, no wrapper.

Sleeps to 0 B

Idle sandboxes cost no memory. Volume intact, back in a second.

Docker or Kubernetes

The same spec local or in a cluster. sbx doctor tells you what a host can do.

Manage from the CLI

sbx ui --connect reaches sandboxes on a VM or cluster and drives them from your terminal.

Snapshot & fork

Seed one database, then fork a hundred. The migration runs a single time.

One committed file

sandbox.json says what a branch needs. Reads with --json for agents.

one screen

Watch every sandbox on one screen.

sbx ui — cpu and memory against each service's limit, a live trace of where it's been, and one keystroke to wake, sleep, cap or port-forward. Local, or over --connect to a deployment.

sbx ui — 2 sandboxes · 17 of 17 services awake · sbx 3.16g of 7.74g

SANDBOXSERVICESTATECPUMEMORYADDRESS

analytics-agentclickhouseAWAKE45mc573m127.0.0.1:20021

analytics-agentmysqlAWAKE47mc158m127.0.0.1:20030

analytics-agentgatewayAWAKE19mc41m127.0.0.1:20027

analytics-agentredisAWAKE28mc4m127.0.0.1:20033

checkout-workpostgresasleep--127.0.0.1:20000

checkout-workcacheasleep--127.0.0.1:20001

analytics-agent/clickhouse cpu ▁▂▃▅▇▆▄▅▇▆▄▂▃▅▇▆▄▃▂▁▂▃▅▇▆ peak 0.9c · 1m

one static binary · zero dependencies

Install in one line.

curl -fsSL https://raw.githubusercontent.com/aryanmehrotra/sbx/main/scripts/install.sh | sh

0 B memory when idle 191 ms redis wake laptop, VM or Kubernetes