Daytona is a secure, elastic infrastructure platform for running AI-generated code — providing sub-90ms sandbox creation, stateful agent workspaces, and a self-hostable OSS stack for teams that need full control over where agentic code executes.
Why Daytona
Most AI agent frameworks need a place to run untrusted code. That environment has to spin up fast (agents invoke sandboxes in tight loops), stay consistent across invocations (stateful dependencies), and be isolatable from the host. Daytona addresses this directly: it positions itself as the compute layer beneath orchestrators like OpenHands, Open SWE, and Hermes Agent — all of which list Daytona as a supported sandbox provider.
The platform pivoted from developer environment tooling to AI agent infrastructure in early 2025, relaunched in late April 2025, and crossed $1M ARR within two months. A $24M Series A (FirstMark Capital, February 2026) funds a ground-up rebuild with the AI agent as the primary consumer rather than a human developer.
Architecture
Daytona organizes its stack into three planes:
- Interface plane — SDKs (Python, TypeScript, JavaScript, Go, Ruby, Java), a RESTful NestJS API, a Go CLI, and a web dashboard
- Control plane — orchestrates sandbox lifecycle: creation, snapshot, teardown, and resource allocation
- Compute plane — runner nodes that host sandbox instances using Docker/OCI images, with optional Kata Containers for hardware-level kernel isolation
Each sandbox is a full composable computer: dedicated kernel, filesystem, network stack, and allocated vCPU/RAM/disk. Access modes include SSH, VNC, web terminal, PTY, and Language Server Protocol (LSP) — making sandboxes usable by both code-executing agents and human-in-the-loop workflows.
OSS Deployment
Daytona publishes its entire stack under AGPL-3.0. Self-hosting uses Docker Compose from the docker/ directory in the repo — the same components as the managed cloud:
git clone https://github.com/daytonaio/daytona
cd daytona/docker
docker compose up -d
For teams with existing compute, the customer-managed compute (BYOC) option runs Daytona's control plane while pointing runners at your own machines or VMs. This covers regulated environments (VPCs, air-gapped networks) — Athena Intelligence is a documented example of deploying agents across regulated VPCs using this model.
Capabilities
- Startup latency: ~27–90ms (Docker container model; faster than Firecracker-based competitors)
- Stateful snapshots: persist pip installs, file state, and environment between agent invocations
- MCP server integration: sandboxes expose MCP endpoints for agent tooling
- Computer Use: Linux, Windows, and macOS virtual desktop environments for GUI agent workflows
- GPU support: configurable GPU allocation for inference-heavy workloads
- Real-time streaming: live stdout/stderr output during code execution
Daytona vs E2B
The main alternative is E2B, which uses Firecracker microVMs (hardware-level isolation, ~150ms cold start) and reports 88% of Fortune 100 companies as signed up. Daytona trades the stronger isolation guarantee for faster startup and stateful persistence. E2B charges a $150/month Pro base fee; Daytona bills purely on usage.
Choose E2B when maximum isolation (dedicated kernel) matters more than startup speed. Choose Daytona when you need sub-100ms agent loops, stateful environments, or full self-hosting.
Key Characteristics
| Property | Value |
|---|---|
| Interface | SDK (Python, TypeScript, JS, Go, Ruby, Java), CLI, REST API, Web UI |
| Provider | Daytona |
| License | AGPL-3.0 |
| Pricing | Usage-based (managed); self-hostable (OSS) |
| Sandbox | Docker/OCI containers; optional Kata Containers for enhanced isolation |
| GitHub | daytonaio/daytona |
| Website | daytona.io |
| Docs | daytona.io/docs |