daytonaio/daytona

Secure, elastic infrastructure for running AI-generated code — sandboxes for agent-produced code execution.

What it is

A platform for running untrusted / AI-generated code in isolated sandboxes. Aimed at AI-agent products that need to safely execute model output without exposing the host: spin up an ephemeral container, run the code, capture results, tear down. Pivoted from a developer-environment focus to AI sandboxing as the market matured.

Key features

Per-request sandbox provisioning with strong isolation.
Multi-language runtime support — Python, JavaScript, shell, more.
Auto-cleanup of completed sandboxes.
SDK + REST API for integration.
Self-hostable or use Daytona Cloud.
AGPL-3.0 licensed.

Tech stack

TypeScript primary on the orchestration layer.
Container-based sandboxing (likely Docker / Firecracker / similar).

When to reach for it

You're building an AI agent that generates and runs code, and need a sandbox layer.
You're shipping a code-interpreter feature in a chat product.
You're operating an LLM-based developer tool that needs safe execution.

When not to reach for it

You need vendor-supported with SLAs at enterprise scale.
You can run trusted code in your own infra without isolation.
AGPL-3.0 is incompatible with your commercial license model — verify.

Maturity signal

73k stars, 5.6k forks, AGPL-3.0, actively maintained. Pivoted positioning from dev-env to AI-sandbox. Open-issues count of 430.

Alternatives

E2B (e2b.dev) — comparable AI-sandbox commercial / OSS product.
Modal — code-execution-as-a-service.
Custom Firecracker / gVisor + your own orchestration.

Wiki: daytonaio/daytona