Preview · Firecracker microVMs on call

A Firecracker box for every agent run.

Each task gets a Linux microVM of its own — real kernel, root, Docker and a disk that stays put. Drive it from the CLI or API, snapshot the state that works, and fork it when you want more than one shot.

Start free Read the docs

01 create ✓ done

moil create --name run-42✓ run-42 is up  firecracker · 4 vCPU · 8 GiB  /data mounted (20 GiB)

02 exec ✓ done

moil exec run-42 -- npm test  Test Files  18 passed       Tests  214 passed✓ exit 0

03 snapshot · fork ✓ done

moil snapshot run-42 --name green✓ saved greenmoil fork run-42 --count 3✓ run-42-1 · run-42-2 · run-42-3

Or grab the CLI preview · not yet published

npx moil@latest

Isolation first

Untrusted code gets a VM boundary, not a shared kernel.

Agents install things, run things and occasionally break things. Moil assumes that, and gives each run walls that are actually walls.

// what every box getsfirecracker · kvm · per-run

A Firecracker microVM per run

Every box boots as its own KVM guest. Agents, tenants and runs never share a kernel.
Root inside, nothing outside

Hand the agent sudo, systemd and native Docker. The blast radius ends at the box.
Keys scoped to one workspace

API keys carry only the scopes you grant and stop at their workspace. Revoke one, the rest keep working.
Secrets that never land on disk

Credentials are added at the egress proxy, for hosts you allow. Code in the box only sees a placeholder.

Lifecycle & snapshots

Boxes that keep their place — and can split in two.

Long agent runs stall, wait on humans and wander off course. Pause them without losing the thread, and branch them before a risky step.

run-42 · running

moil resume run-42

Pick a state, or let it cycle.

What each box state keeps
Property	Running	Paused	Stopped
Files & /data	kept	kept	kept
Memory & processes	live	frozen, kept	gone
Comes back with	—	resume, mid-thought	a fresh boot
Metered (preview)	compute + disk	disk	disk

Pause mid-thought. Resume mid-thought.

Paused boxes keep memory and processes; stopped boxes keep their files.

Idle auto-pause you can tune per box
A persistent /data disk across restarts
New boxes that boot from a snapshot
Wake on CLI, API, SSH or an inbound request

Fork the run before it goes sideways.

Copy the box on its host, then let three attempts race.

Root disk and /data copied in place
Same template, same size, new name
The source box keeps running untouched
Keep the winner, delete the rest

Give the agent a screen. Keep a hand on the mouse.

A Linux desktop and a real browser, driven over the SDK.

Screenshots, clicks and keystrokes as API calls
Chromium that keeps its sign-ins between runs
Take the wheel, then hand it back
Share a watch-only link to the desktop

Your CI is one `runs-on` line away.

Swap the label, keep the workflow you already have.

How CI runners work

A clean box per job, gone when it ends
Ubuntu 24.04 with the usual toolchains
Sizes from 2 to 32 vCPU
No GitHub App? Call the CLI from any CI

.github/workflows/test.yml

name: teston: [push, pull_request]jobs:  test:    runs-on: ubuntu-24.04    runs-on: moil-medium    steps:      - uses: actions/checkout@v4      - run: npm ci && npm test