Prateek Bisht
All work

claude-code-microVM

agents · agent sandboxes

chatsmicroVMs · KVM · ~0.3 s bootsessionschat 1vm-1 · claude codechat 2vm-2 · claude codechat 3vm-3 · claude codeS3

Every chat gets its own coding agent, inside its own throwaway microVM.

What it is

Agents are most useful when they can run anything, and safest when nothing they run can escape. Each conversation boots a KVM microVM on EC2 through a self-hosted CubeSandbox, so the agent gets full tool access and the blast radius stays one disposable VM.

Files and session state live on a per-chat S3 volume. Idle VMs are torn down and the chat picks up later on a fresh one, which lets a single box carry several sessions at once. Teammates plug in their own MCP tools with one-click OAuth and watch the agent work live in the browser.

How it works

  1. 1BootA microVM starts in about 0.3 s, for that chat only.
  2. 2RunClaude Code works with full tool access, isolated from the host and other chats.
  3. 3ParkAfter 5 idle minutes the VM is destroyed. Files stay on S3.
  4. 4ResumeThe next message restores the session on a fresh VM.

In numbers

  • Cold start~0.3 s
  • Idle teardown5 min
  • Sessions per box3–4
  • OAuth providers verified10

Built with

  • Bun
  • TypeScript
  • React
  • KVM
  • AWS EC2
  • S3
  • MCP

Links

GitHub

Next

plan-out →