# Daytona

> Sub-90ms agent sandboxes — isolated computers with snapshots, volumes, Git and LSP tools, on Linux, Windows, or Android; AGPL self-host or managed cloud.

Daytona pivoted from dev-environment manager to agent infrastructure and found its market: sandboxes that start in under 90ms — isolated computers with dedicated kernel, filesystem, and network, lifecycle primitives, shared volumes, and agent tools — on Linux, Windows, or Android, with GPUs available. AGPL-3.0 self-hostable; cloud is usage-billed with signup credits.

Website: https://www.daytona.io

Daytona is the category's speed-and-breadth play, and one of 2026's cleaner pivot stories: the dev-environment manager rebuilt itself as **infrastructure for agent code execution** — "give every agent a computer" — and the market answered (a FirstMark-led $24M Series A in February 2026, with LangChain among the customers).

## Highlights

- **Sub-90ms sandbox creation** — fast enough that agents treat computers as disposable per-step resources, not provisioned assets.
- **Real isolation** — dedicated kernel, filesystem, and network stack per sandbox, with configurable vCPU/RAM/disk and GPU options.
- **Lifecycle primitives** — start, stop, pause, **snapshot**; stateful sandboxes persist across runs, and volumes share data between them.
- **Multi-OS** — Linux by default, with Windows and Android sandboxes (priced per vCPU-hour) — the unusual capability for testing and automation beyond the Linux monoculture.
- **Agent-shaped tooling** — process exec, filesystem ops, Git operations, and LSP support exposed through SDKs in Python, TypeScript, Ruby, and Go.
- **Three deployment modes** — managed cloud, fully self-hosted AGPL stack, or hybrid control-plane over customer-managed compute.

## In an AI-assisted workflow

```bash
pip install daytona      # or: npm install @daytona/sdk
# sandbox = daytona.create(); sandbox.process.exec("python analyze.py")
```

The fit: agent systems that churn through many short-lived executions (where 90ms vs seconds compounds), need Windows/Android targets, or have compliance reasons to self-host the whole stack.

> [!NOTE]
> Two npm scopes exist from the transition — legacy `@daytonaio/sdk` and current `@daytona/sdk` (both published, README uses the new one). And per-second billing means idle sandboxes cost money until stopped: wire cleanup into the agent's lifecycle.

## Good to know

AGPL-3.0 (copyleft applies to self-host modifications), releases in fast lockstep across PyPI/npm. The honest caveat repeated from our research: the ~72k GitHub stars substantially predate the pivot — judge adoption by the 2026 product, not the counter. Category trade-offs versus [E2B](/tools/e2b), [Modal](/tools/modal), and [Vercel Sandbox](/tools/vercel-sandbox): [Sandboxing AI-Generated Code](/guides/advanced/sandboxing-ai-generated-code).

---

_Source: https://agentscamp.com/tools/daytona — Tool on AgentsCamp._