Reverse-Engineering Claude Code Web's Hidden Firecracker MicroVM

Researchers investigating Claude Code Web's runtime discovered it runs inside Firecracker microVMs, the same lightweight virtualization technology powering AWS Lambda and Fargate. Using standard Linux tools like strace and dmesg from within an active session, they traced ACPI signatures, kernel command-line flags, and a minimal PID 1 process tree back to Firecracker's known implementation details.

The investigation revealed a snapshot-based session architecture: VM templates are pre-booted, frozen, and later restored via hot-swapped block devices, with an ext4 rootfs and squashfs overlays for the Claude Code CLI and environment runner mounted at restore time. A custom Rust/tokio binary called process_api serves as both init and a WebSocket-based remote process supervisor, exposing structured process creation, I/O, and lifecycle APIs alongside an HTTP control plane for mounting rootfs and managing container identity.

Perhaps the most notable finding was an unstripped 27MB Go binary, environment-runner, containing full debug symbols and referencing an internal Anthropic module path, hinting at an undocumented internal PaaS used to orchestrate these sandboxed coding sessions.

For engineers building sandboxed execution environments, agent runtimes, or Firecracker-based platforms, the findings offer a rare, concrete look at how a major AI lab implements ephemeral, snapshot-restored compute for coding agents, including security steps like CRNG reseeding, capability dropping, and page-cache clearing between sessions.