Run real code. Safely.
Shuttle gives an AI its own VM, isolated by hardware, where it writes, compiles and runs real code in six languages — without ever touching your machine. Ready in 122 milliseconds.
The problem
A container shares the kernel.
AI agents are only useful when they run the code they write — build, test, verify. But running AI-generated code on your machine, or even in a container, is real risk: a container shares the kernel, and one escape means a compromised host.
The solution
A capsule isolated by hardware.
Each execution runs in a KVM capsule — a full VM, isolated by hardware. An escape can't reach the host because the boundary is the processor, not the kernel. And thanks to fork-COW cloning, that capsule is born in milliseconds: total isolation, no speed tax.
Proof · measured on bare metal
Numbers, not adjectives.
NodePythonGoRust.NETJava
Why it's different
Four things most sandboxes can't say.
Hardware, not container
KVM is a real boundary — the processor, not a shared kernel.
Fast despite isolation
Fork-COW makes a full VM ready in milliseconds.
Real builds, not snippets
A real project with dependencies, in six languages.
Auditable TCB
The whole isolation supervisor is 589 lines of C.
Works with
One MCP server. Every client.
Shuttle is MCP-native: a single server (mc mcp) plugs into the tools you already use. Configure once.
Claude Code
Codex
VS Code · Copilot
Capacity scales with the host. Sustained throughput and concurrent-capsule limits are still being measured under load — we'll publish them when the numbers are real, not before.