Self-hosted sandboxes for AI coding agents, so they stop needing a babysitter
AI coding agents are good enough to trust with real implementation work, but they need constant supervision: approving file writes, answering questions, babysitting progress across multiple terminals. L’Atelier gives each agent its own disposable, isolated sandbox and lets you dispatch work like tickets on a kanban board, then walk away.
Each sandbox is a full, hardware-isolated environment (VS Code, an AI coding agent (OpenCode), a browser, SSH) that boots in seconds and comes with a complete dev environment out of the box. You describe a task from a dashboard or directly from Slack, the system clones the right repo, spins up a sandbox from a pre-warmed snapshot, and starts the agent. You get progress updates, answer questions when the agent is stuck, and review a diff when it’s done, all from a phone if needed.
Cloud sandbox vendors (Codespaces-style, or Ramp’s internally-built “Inspect”) solve this well, but are either vendor-locked, per-seat priced, or require infrastructure most teams shouldn’t pay for. L’Atelier is the self-hosted version: one bare-metal server, no external dependencies, real VM-grade isolation instead of shared-kernel containers, because an AI agent with root access is one kernel exploit away from owning the host if it’s “just” a container.
I designed and built the entire system solo: orchestrator, Rust in-VM agent, Slack integration, Helm chart. It’s not a personal toy: it’s infrastructure the team uses daily at Frak to dispatch and review AI-assisted engineering work, and it continues to evolve (multi-node support, warm pools, SSH proxy are next).
The orchestrator gets all the glory, but it's the proxy, the cache, and the registry that make the system actually usable.
An AI sandbox orchestrator needs more than just an orchestrator. Here's how CLIProxy, shared binary volumes, Verdaccio, and a single Helm chart turn L'Atelier from a VM launcher into a dev platform.
Migrating from a custom Firecracker orchestrator to Kata Containers and k3s. Better portability, less maintenance, and a lot of deleted code.
L'Atelier just went through its biggest architectural shift yet. I swapped a custom bare-metal Firecracker orchestrator for Kubernetes and Kata Containers: deleted 8,702 lines of code, gained portability, and traded sub-second boots for a system I can actually maintain.
Three weeks of daily use revealed the real bottleneck. The dashboard was still one too many clicks away.
L'Atelier now dispatches AI coding agents directly from Slack, powered by an MCP server and a system sandbox running an AI dispatcher agent. Here's how the architecture evolved.
Building a self-hosted Firecracker microVM orchestrator because AI coding agents shouldn't need you watching.
AI coding agents are powerful, but they need you. Approving writes, answering questions, babysitting progress. L'Atelier gives each agent its own VM sandbox, lets you dispatch tasks like a kanban board, and review results from your phone. I built it right before a ski vacation. It worked.
Curious about the technical details, or have a similar challenge? Reach out.
Book a call