Sub-90ms sandboxes for AI-generated code

What it does

Daytona is an open-source runtime that provisions fully isolated sandboxes—complete with dedicated kernels, filesystems, network stacks, and allocated compute—for executing AI-generated code and running agent workflows. Agents and developers interact with them through multi-language SDKs, a REST API, and a CLI to manage lifecycle, filesystem, process execution, and runtime configuration. The platform also snapshots sandbox state for persistence across sessions, and offers human-facing access via a web dashboard, SSH, VNC, and a web terminal.

The interesting bit

Each sandbox is pitched as a “full composable computer” with its own kernel, yet the platform builds on OCI/Docker compatibility and promises sub-90ms cold starts. That combination—dedicated-kernel isolation with container-image ergonomics—is the central engineering bet.

Key highlights

• Multi-plane architecture splitting interface, control, and compute into distinct components including a NestJS API, Go CLI, and dedicated runner nodes.
• OpenAPI-generated SDKs for Python, TypeScript, Ruby, Go, and Java, plus toolbox API clients.
• Deployment flexibility: managed SaaS, self-hosted Docker Compose stack, or hybrid with customer-managed compute.
• Agent-centric tooling including LSP support, an MCP server, pseudo-terminals, log streaming, and Git operations.
• Enterprise controls: organizations, API keys, audit logs, billing, network limits, and OpenTelemetry tracing.

Caveats

• The README never names the hypervisor or isolation mechanism behind the “dedicated kernel” claim, and the “under 90ms” figure lacks context on hardware or image size.
• The documentation orients new users toward the managed service at app.daytona.io first; the open-source stack reads as a bring-your-own-infrastructure alternative.

Verdict

Evaluate it if you are building agentic workflows that need fast, isolated, and stateful code execution with enterprise guardrails. Skip it if you just need a simple local container runner or want to avoid platform overhead.