SandboxesOpen beta

A sandbox is a step.

Create microVMs inside your Inngest functions. They retry, resume after a crash, and show up in the trace like every other step. Watch one run:

Get started Read the docs

Full playground

inngest/create-sandbox.ts


export const createSandbox =
  inngest.createFunction(
  { id: "create-sandbox" },
  { event: "playground/create.run" },
  async ({ step }) => {
    const created = await step.sandbox.create(
      "create-sandbox",
      { vcpu: 2, memoryMb: 2048 },
    );
    const sandbox = await created.waitUntilRunning(
      "wait-running",
    );
    const result = await sandbox.commands.run(
      "run-command",
      "node --version",
    );
    await sandbox.destroy("destroy-sandbox");
    return result.stdout;
  },
);

Want the controls? Pick a command, inject a failure, and read every step’s input and output.

Open the playground

Use cases

Durable machinesfor every agent.

A sandbox is a step. It retries, it shows up on the trace, and the same flow control you already use applies to the run.

inngest/analyse-data.ts

// One machine per run. Commands are steps.
const sandbox = await step.sandbox.create(
  "create-sandbox",
  {
    name: `analyst-${runId}`,
    vcpu: 1,
    memoryMb: 1024,
  },
);
 
const result = await sandbox.commands.run(
  "run-code",
  ["python3", "-c", code],
);

Code interpreter

Run AI-generated code

If you ask a question, and your agent writes Python to answer it, that script needs a machine. Each question gets its own sandbox: upload the file, run the code, and the output comes back as the next step.

Read the docs

inngest/score-code.ts

// Run each case sandboxed. Score after.
const result = await sandbox.commands.run(
  `case-${i}`,
  'printf %s "$INPUT" | python3 -c "$CODE"',
  {
    environment: {
      CODE: code,
      INPUT: test.stdin,
    },
    timeout: "5s",
  },
);
 
defer("score", {
  function: passRate,
  data: { code, cases },
  experiment: experimentRef,
});

Evals

Score generated code

A new prompt only counts if the code it writes still passes. Run each case in a sandbox, then score in the background so the result is credited to the prompt that wrote it.

Read the docs

inngest/process-dataset.ts

// Install once, snapshot, clone per file.
await builder.commands.run(
  "install",
  "pip install --quiet duckdb",
  { timeout: "5m" },
);
 
const environment = await builder.snapshot(
  "snapshot-environment",
);
 
const worker = await environment.clone(
  `clone-${i}`,
  { name: `dataset-${runId}-${i}` },
);

Custom environments

Heavy jobs, off your servers

A DuckDB pass, a media batch, or a build will swamp the server handling requests. Install the tools once, snapshot the environment, and clone a worker per file. Each file retries on its own.

Read the docs

Why Inngest