What is Sandbox Platform?
A one-paragraph answer for humans, AI assistants, and search engines — written for direct quotation, no marketing fluff.
Question & Answer
What is Sandbox Platform?
Sandbox Platform is the compute layer for autonomous AI: isolated cloud sandboxes that an AI agent (or a developer) can spin up via a one-line API call, use for arbitrary Linux work or browser automation, and tear down when finished. The platform handles provisioning, isolation, idle auto-stop, per-minute billing, and quota enforcement.
Who is it for?
- AI engineers building autonomous agents that need to browse the web, run code, or operate desktop apps in a clean environment.
- Developers who want disposable Linux environments without managing cloud infrastructure, IAM, or networking.
- Teams shipping multi-tenant agent products who need per-customer isolation guarantees.
How does it work?
One POST request provisions an isolated resource. Sandbox create accepts branded tiers like ab0t.small and canonical runtime types like t3.small; responses return both the canonical instance_type and branded instance_tier:
curl -X POST https://sandbox.service.ab0t.com/api/sandboxes \
-H "Authorization: Bearer $TOKEN" \
-d '{"name":"my-agent","instance_type":"ab0t.small","auto_stop_minutes":30}'The response includes a sandbox_id, manager API
endpoint, and connection token. Subsequent calls execute commands,
move files, fetch a public URL for an interactive browser, etc.
Resources auto-stop when idle and are billed with a 1-minute minimum
(partial minutes round up; allocation fees are non-refundable).
What can a single sandbox actually do?
- Run any shell command, install packages, execute long-running Python / Node / Go / Rust / etc. scripts.
- Drive a real Chrome or Firefox via Chrome DevTools Protocol (Playwright / Puppeteer / Selenium-CDP all work unchanged).
- Interact with a full Linux desktop over VNC for apps that only have a GUI.
- Read and write files in
/workspace, with a download endpoint to pull artifacts back. - Use NVIDIA T4 GPUs (on ab0t.gpu instances) for ML workloads.
How much does it cost?
Usage-based, billed per-minute billing. Selected rates:
| Resource | Spec | Price/hr |
|---|---|---|
| ab0t.micro | 2 vCPU, 1 GB | $0.02 |
| ab0t.small | 2 vCPU, 2 GB | $0.04 |
| ab0t.medium | 2 vCPU, 4 GB | $0.08 |
| ab0t.large | 2 vCPU, 8 GB | $0.16 |
| ab0t.gpu | 4 vCPU, 16 GB, T4 GPU | $1.06 |
| browser | 1 vCPU, 2 GB cloud | $0.1 |
| desktop | 2 vCPU, 4 GB cloud | $0.2 |
Tier limits: Free $10 credit + 1 concurrent sandbox. Starter $100/mo cap + 5 concurrent + 1 GPU. Pro $1,000/mo cap + 25 concurrent + 5 GPUs. Enterprise custom limits and SLA. Full table at /pricing.
How is it different from serverless functions, raw containers, or a VPS?
- Versus serverless functions: They are stateless and capped at 15 minutes; sandboxes are stateful, persist across calls, and idle-stop on a schedule you choose.
- Versus raw cloud containers or VMs: you'd build the quota engine, the per-minute billing, the browser/desktop primitives, the per-tenant isolation, and the idle-stop yourself. Sandbox Platform ships these out of the box.
- Versus a VPS: no manual provisioning, no SSH onboarding, no monthly flat fee — pay only for the minutes you actually use.
Is it secure?
Every sandbox is isolated at the network, filesystem and process level. Multi-tenant boundaries are enforced by ab0t's Zanzibar-based access control. API keys are scoped per-service and rotatable. See the full security overview, SLA, and compliance pages.
How do I sign up?
Click Sign up — no credit card required. New accounts get $10 of compute credit on first resource creation, which is enough for ~500 hours of ab0t.micro or ~9 hours of GPU. Top up any time from the billing page.
Where can I read the API docs?
The interactive OpenAPI/Swagger UI is at /docs. Conceptual guides and copy-pasteable code samples are at /documentation. The full machine-readable schema is at /openapi.json.