Sandboxes

An isolated computer for every agent run

Run code, process files and call your systems, with the Adapt company agent or your own. What a model writes never runs inside your app.

q3-board-pack

Medium · 4 vCPU · 4 GiB

Parked

$ python report.py

✓ Loaded 3 exports

✓ Joined accounts and invoices

✓ Found 4 accounts behind plan

✓ Charted revenue by region

✓ Wrote report.pdf

Woke in–
Billed
Cost

Use cases

What you can build with sandboxes.

  • Run untrusted code

    Execute what a model writes in a machine that can’t touch your app.

  • Process large files

    Load exports, transform them, return a report or a dataset.

  • Long-running jobs

    Hand off work that takes minutes, and get the result when it’s done.

How it works

Prepare a report from several large exports.

The agent coordinates the task while the sandbox handles the data processing.

  1. Load the files

    Make the approved source files and tools available to the job.

  2. Run the transformation

    Run code to normalize inputs, join records and calculate the metrics.

  3. Return the result

    Save the report and return a summary with its location, ready for review.

sandboxes.ts
import { Adapt } from '@adapt/sdk'

const box = await new Adapt()
  .sandboxes.create({ size: 'medium' })

await box.files.upload('./q3-exports')
await box.run('python report.py')
await box.files.download('report.pdf')
Illustrative API · final SDK may differ

Pricing

What it costs to run.

Metered per second while it runs, and a parked sandbox costs nothing.

  • $0.0351per vCPU-hour, billed only while running
  • $0.0112per GiB-hour of memory
Sandbox sizes
SizevCPUMemory
Small22 GiB
Small, more memory24 GiB
Medium44 GiB
Medium, more memory816 GiB
Large1616 GiB
XLarge3232 GiB

Works with

Use it with the rest of Adapt.

The company agent combines them with your tools, knowledge and team’s apps.

Start building with Sandboxes