disposable linux · no accounts

A Linux box
with a very
short life.

throwbox rents Linux machines by the session. Pick a length, pay once in USDC on Base, run your thing. The lease ends and the machine is gone, files and all. There is no suspend and no resume, so nothing you write outlives the lease.

MIT source · hosted beta at api.throwbox.dev

SANDBOX · ONE SESSION NO. 6174-X402

DAY PASS

one (1) linux machine

cpu
1 vCPU
memory
1 GiB
lease
300 s
fare
x402

Non-transferable. Non-persistent. Void after 300 seconds.

ONE AGENT SBX-05
127 ms
warm session ready, p50
53 ms
exec round trip, p50
182 MiB
idle guest memory
0
accounts, keys, signups

Two-node Kata cluster, control 0.1.16, 2026-11-09, n=50. 0.1.22 measured 292 to 354 ms server ready across two runs, so this is a baseline, not a promise. Method and caveats.

the idea

Rent the machine,
not the relationship.

No signup, no dashboard, no key rotation.

01

No account.

The payment is the identity. A retried payment returns the same session instead of billing twice.

one payment → one session

02

No peeking.

Commands and files are sealed to the guest's key. The API, controller and relay route ciphertext only, and hold no key to any of it.

ciphertext in transit · Kata microVM · boundary tested · escape untested

03

No leftovers.

Three things end the lease: the guest watchdog, the controller deadline, Kubernetes. The pod and its workspace go with it, and nothing is written somewhere that outlives them.

delete is idempotent · no suspend, no resume

the life

Born, useful,
gone.

The lease is the only thing that can keep it alive. There is no suspend and no resume here: the workspace dies with the lease, and the next session starts empty.

  1. 0 ms

    Knock

    Post a spec and a public key. Get a price for that spec.

    POST /v1/sessions → 402
  2. provisioning

    Wake

    Cold boot, or a warm pool if one is running. Capability and session come back once.

    201 · READY
  3. 300 s

    Work

    Sealed commands, files under /workspace, one foreground operation at a time. Background processes die when the command returns.

    exec · file.put · http
  4. 0 s

    Gone

    Lease ends, pod deleted, workspace purged. There is nothing to clean up.

    DELETE → 200

three handles

Pick the smallest handle.

Use the CLI or the SDK. Don't hand-roll the envelopes.

for your agent

Now hand the robot
the manual.

Install, one-shot path, HTTP flow, limits, failure modes. One file.

  • Why hand-rolled envelopes fail
  • How to destroy in a finally block
  • What outcome_unknown means
  • What a 402 or a 429 means
  • What isolation does not promise
Open llms.txt

handoff.txt READ FIRST
# tape this to your agent
Read {origin}/llms.txt first.
Hosted API: https://api.throwbox.dev

Use the CLI or @x402s/sdk.
Do not hand-roll the sealed envelopes.

Prefer sandbox run --json.
Destroy in a finally block.
Keep files under 8 MiB.
Never inline secrets in a command.

If the result says outcome_unknown,
the command may have run. Do not
blindly retry a side effect.
curl -s {origin}/llms.txt

on your laptop

Two minutes to your own.

Same API, relay and guest as a real deployment.

01

Local cluster

make up            # kind + cilium + images + deploy
make demo-session  # pay, create, exec, upload, download
make down

Needs Docker, kind, kubectl, the Cilium CLI, Node 24+ and Go 1.26+. Local clusters use runc, not microVMs. API on http://localhost:8080; GET /healthz is the cheapest check.

02

No cluster

npm ci
make verify  # typecheck, structure lint, unit + property
make test    # adds integration against the guest binary

Most of the system is testable in under a minute. make security-test needs a cluster and runs the packet-level isolation suite.

Hosted beta, one operator, one region, no SLA. Payment is real USDC on Base. One control-plane node and one Postgres primary carry the control plane, so losing that node is an outage and no failover has been measured. Sessions are ephemeral: no suspend, no resume. Isolation is a Kata microVM, tested for the boundary properties in the suite; that is not a hardware root of trust, and no escape has been attempted. What the isolation run proved.

the rules

Four things, plainly.

You pay per session, we run the machine, and neither of us wants a contract to read.

Acceptable use

Use a session for work you have the right to do. Not allowed: attacking or scanning systems you do not own, spam, crypto mining, illegal content, escaping the sandbox, or reaching another session, the cluster network or its metadata service.

Report abuse to abuse@throwbox.dev. That address reaches the operator, because the operator is the whole company.

Money

One payment, one session, in USDC on Base, settled through the Coinbase CDP facilitator. There is no account to top up and no balance to spend later: what you pay covers the session you asked for, and the lease ends when it ends. Free access is by invitation only.

If a session never reaches ready, or we end it because the relay could not recover, the payment is refunded in full to the paying wallet. A session you destroy, or one that simply expires after it was ready, is not refunded: the time was delivered.

Refunds are queued in the payment ledger and sent by a separate worker, so they are not instant. If one cannot be sent automatically it stays queued and an operator resolves it.

What we record

Session metadata (tier, limits, timings, state), the paying wallet address, and the payment ledger. Commands, file contents and output stay ciphertext: the API, controller and relay hold no key.

The guest sees whatever you run, and nothing you write is kept: the workspace dies with the lease. Checkpoint storage is plaintext, which is why there is no suspend. This page sets no cookies and runs no analytics.

What we do not promise

No SLA and no uptime target. No hardware attestation, so the provider could run a different guest. The boundary is tested for the properties in the suite; no escape has been attempted, and the review is not independent.

One operator, one region, one control-plane node, one Postgres primary. If that node is lost, the control plane is down until it comes back.

Clone it. It won't remember you.

Bring a cluster, read the protocol, or point a client at api.throwbox.dev.

github.com/ejackson-vp/sandbox