Give your agent
a real computer.
Isolated Linux microVMs behind a simple API — for agent evals and always-on production runs. Start one in seconds, run code you didn't write, keep what it produced, and stop paying the moment it's done. Hosted entirely in the EU.
Evals and production runs, on the same fleet
A sandbox is the same primitive whether it lives for ninety seconds or nine hours. Run the bursty test sessions and the long-lived production ones the same way — the API doesn't change, only how long you keep the machine.
Evals & bursts
Fan out hundreds of short-lived sandboxes for a test suite or a benchmark, run model-written code in parallel, collect the results, tear them all down. You pay for the seconds they ran and nothing for the idle in between.
Production agent runs
Keep a sandbox alive for a whole production session — an autonomous agent acting on real customer data. Checkpoint to pause and resume, with idle auto-pause and hard spend ceilings so a forgotten agent can't run up an open-ended bill.
A machine with a beginning and an end
Sandboxes are designed to be created, used, and destroyed — the whole lifecycle is a handful of HTTP calls, and billing follows the state.
- template
Define the environment once — packages, setup commands, CPU, and memory. Versions are immutable.
- running
Create a sandbox from a template version. It boots in seconds, ready for streaming processes and files.
- checkpoint
Snapshot the disk mid-task. Restore it into a fresh sandbox to pause, branch, or recover a session.
- stopped
Stop to release compute and pay only for the retained disk. Connect again for instant resume.
- destroyed
Done with the work? Destroy the sandbox and stop paying entirely. Nothing lingers.
Everything an agent workflow needs
No SDK required, no session daemons to run — a Watasu API key and an HTTP client are enough.
Streaming processes and files
Run commands and PTYs over live sockets, watch file changes, use signed file URLs, and keep regular file reads and writes as plain HTTP.
Isolated by default
Each sandbox is walled off from your apps and from every other sandbox. Code your agent just wrote can touch exactly what you handed it — nothing else.
Preview URLs
Expose a port and get an HTTPS URL at a `*.sandbox.watasuhost.com` address. Show a running result without deploying anything.
Reusable templates
Capture packages and setup once; every sandbox starts from a pre-built, versioned image. No per-task install steps eating your seconds.
Checkpoints
Named disk snapshots you can create, restore, and delete. Long agent sessions become pausable, branchable, and recoverable.
Stop and instant resume
A stopped sandbox keeps its disk for cents per month. Reconnect and it resumes with the same files, same sandbox id, in seconds.
Built for code nobody reviewed
If a program was written seconds ago — by a model, by a user, by a pipeline — it shouldn't run next to anything you care about.
Agents that need a computer
Give every agent session its own machine: a real shell, a real filesystem, a preview URL. Create on task start, checkpoint along the way, destroy on completion.
Generated code, contained
Run model-written programs behind an API call instead of on your own infrastructure. Whatever the code does, it does inside the sandbox — and only there.
Code execution as a feature
Notebooks, plugins, user-defined automations, CI-style jobs: per-user isolation with per-second costs that track real usage instead of reserved servers.
The compliance deadlines have already passed
If your agents touch EU data, the regime governing where they run is in force now — and a US provider opening an EU region doesn't make the exposure go away. If your customers or your regulator demand EU residency, this is the rare option that can actually meet it.
The rules are already in force
The EU AI Act has applied to prohibited practices since 2 February 2025 and to general-purpose AI models since 2 August 2025. DORA has bound financial entities to manage their ICT third-party dependencies since 17 January 2025. The EU Data Act added third-country-access safeguards on 12 September 2025. These deadlines are behind you, not ahead.
A US region does not fix the exposure
Under the US CLOUD Act, US authorities can compel any US-controlled provider to produce your data wherever it is stored. Transfers still ride on Standard Contractual Clauses and a Transfer Impact Assessment under Schrems II. An EU region opened by a US company does not change who can be served the warrant.
Watasu keeps it in the EU
Your sandboxes run in EU data centers and never cross the Atlantic. Watasu is not a US company and is not subject to the US CLOUD Act, and a Data Processing Agreement is available from day one. For a team whose customers or regulator demand residency, that is the difference between shipping and not.
References: Regulation (EU) 2024/1689 (AI Act) · Regulation (EU) 2022/2554 (DORA) · Regulation (EU) 2023/2854 (Data Act) · GDPR Ch. V and CJEU C-311/18 (Schrems II) · 18 U.S.C. §2713 (US CLOUD Act). See our security & compliance posture.
Pay for seconds, not servers
Runtime is billed per second for the vCPU and memory you request. Storage for stopped disks, checkpoints, and templates is prorated monthly.
| Runtime — billed per second | |
|---|---|
| vCPU | €0.000010/vCPU·s |
| Memory | €0.000002/GiB·s |
| Storage — per GiB, prorated monthly | |
|---|---|
| Stopped sandbox disk | €0.08/GiB·mo |
| Checkpoints | €0.02/GiB·mo |
| Templates | €0.02/GiB·mo |
A 2 vCPU / 1 GiB sandbox costs about €0.08 per hour while it runs — a 90-second job costs a fifth of a cent. And because Watasu is prepaid, a runaway agent can only spend your balance — never run up an open-ended bill. Read the sandbox billing docs.
Your first sandbox is a curl away.
Create an API key in the dashboard, build a template, and start running code that arrives faster than you can review it — safely.