A sandbox API, not a compute platform
Modal and this service get compared because both end with your code running on someone else's machine. They are answers to different questions. Modal is for running your code somewhere bigger; this is for running code you did not write somewhere disposable.
Which shape is your problem?
| If you are trying to… | The better fit |
|---|---|
| Run a Python function on 32 cores or a GPU, at scale, on a schedule | A compute platform. Modal is very good at this and we do not compete. |
| Execute code an LLM just wrote, before you trust it | A disposable machine. That is this. |
| Let your users submit code and grade it | A disposable machine, one per submission. |
| Give an agent a shell it can break without consequence | A disposable machine. |
| Build and cache a custom container image, then invoke it | A compute platform. We have no image building. |
| Run a build or a test suite for an untrusted repository | Either — but you will want the shell. |
The difference that shows up in the code
With a function platform you describe an environment ahead of time and then invoke it. That is
exactly right when you know what will run. It is awkward when you do not: an agent that decides
mid-run it needs ffmpeg is not a code path you can decorate in advance.
Here there is nothing to describe. You get a machine with Python 3.12, Node 22, git and build-essential on it, and a shell. Anything else, the code installs for itself at the time. That flexibility is the whole point, and it costs you image caching — the trade is deliberate and it is the wrong trade for a workload you control.
What you get for $0.09 an hour
| Size | vCPU | Memory | Disk | Price |
|---|---|---|---|---|
small | 2 | 4 GB | 40 GB | $0.09/hour |
medium | 4 | 8 GB | 80 GB | $0.28/hour |
large | 8 | 16 GB | 160 GB | $0.55/hour |
Billed by the second from prepaid credit, with no plan fee, no seats and no minimum. Provisioning time is not billed. A machine takes about half a minute to be ready, because it is a real VM booting rather than a container starting — the measured figures are in the limits documentation.
Where we would send you elsewhere
If you need GPUs, custom images, persistent volumes, a public URL for something running inside the sandbox, sandboxes living longer than 24 hours, or an official SDK, this service does not have them and Modal does. We would rather say so here than have you discover it after an afternoon of integration work.
What we do have is a smaller surface: create, exec, files, extend, destroy. It is documented in an OpenAPI spec you can generate a client from, there is no plan fee, and the credit is prepaid so the most you can lose is what you topped up.
Questions
Is this a drop-in replacement for Modal?
No, and it would be dishonest to imply it. Modal is a compute platform: you decorate Python functions, it builds images, ships your code, scales it out, schedules it, and gives you GPUs, volumes and web endpoints. This is one API that creates a Linux machine, runs shell commands in it, moves files, and destroys it. If your problem is "run my Python somewhere bigger", Modal is built for that and this is not. If your problem is "run somebody else's code somewhere safe", the shapes reverse.
Why would I choose a plain machine over decorated functions?
Because the code you are running is not yours to decorate. An agent that writes a shell script, a user who submits a solution, a build that has to run make — none of that fits a function signature. You want a filesystem, a shell, and the ability to run whatever appears, without a build step between having the code and running it. That is the entire product here.
What does Modal have that you do not?
GPUs, published SDKs, container image building with layer caching, persistent volumes, scheduled jobs, web endpoints for functions, and a substantially larger company behind it. We have no GPUs, no SDK packages yet, no custom images, no persistent storage, no public URL for anything running inside a sandbox, and a 24 hours ceiling on how long one lives. Those are real gaps, not framing.
How does the cost compare?
They are not directly comparable, because Modal bills CPU and memory separately by the second while running and charges nothing between invocations, which suits bursty function workloads very well. Here you pay for a machine for as long as it exists — $0.09 an hour for 2 vCPU and 4 GB, billed by the second with no minimum. For a sandbox that lives minutes at a time the arithmetic is simple and usually favourable; for a function invoked ten thousand times a day for 200 ms each, a serverless model wins outright and you should use one.
Can I use both?
People do, and it is a sensible split: the agent framework and the model calls run wherever your application already lives, and the untrusted execution goes somewhere disposable. Nothing here expects to be your compute platform.
Try it against your own workload
$5 of credit on signup — about 56 sandbox-hours — and no card required.