open source

The local-first embedded runtime for AI agents.

Real microVMs, as a library. No daemon, no cloud, no root - on your laptop, your cluster, your customer's machine.

[ Read_the_docs ]
CLAUDECODEXGEMINIHERMES

features

Four things containers can't do.
Clouds shouldn't have to.

$ box.boot()

Microsecond microVMs

Real Linux kernel per box. Own kernel, own memory, own everything. Hardware isolation at container speed.

<50ms
box.resume()

State that survives

Install once. Edit once. Stop and resume tomorrow - packages, files, env still there.

T+1w
snapshot -> fork

Fork like a branch

Snapshot the full state and branch it. Explore three strategies in parallel. Roll back risky ops instantly.

3x
pip install boxlite

One library. Every platform.

Same pip install on macOS, Linux, and Windows. Same code on your laptop and your production cluster.

KVM

quickstart

Three lines. Any language.
A real VM.

Not a sandbox. Not a container. A real Linux VM, called from a function.
01import asyncio02import boxlite0304async def main():05    async with boxlite.SimpleBox(image="python:slim") as box:06        result = await box.exec("python", "-c", "print(2 + 2)")07        print(result.stdout)  # "4"0809        await box.exec("pip", "install", "numpy")10        result = await box.exec(11            "python", "-c",12            "import numpy as np; print(np.random.rand(3))"13        )14        print(result.stdout)1516asyncio.run(main())

runtime

A library on top.A microVM underneath.

Real microVMs. Your machine.

Containers are embeddable but share a kernel. MicroVMs isolate properly but need a daemon. BoxLite eliminates the tradeoff.

Rust core, native bindings. No daemon. Pluggable hypervisor: KVM on Linux, Hypervisor.framework on macOS, WSL2 on Windows.

use cases

Built for the agentspeople actually ship.

FIG 01
One runtime. Every shape.

All use cases →

RUN CLAUDE CODE IN A BOX

An autonomous coding agent confined to a microVM.

The agent gets root inside its own kernel and nothing at all outside the box.

BUILD A CODE INTERPRETER FOR YOUR LLM

A model writes Python; it executes in a disposable microVM.

Generated code never touches your host, and the box is thrown away when it is done.

REVIEW UNTRUSTED PULL REQUESTS

Tests and an automated review of code you have not read.

Clone the PR into a box, run the suite, then read the diff - with no access to your machine.

BUILD A DATA ANALYSIS AGENT

CSV in, model-written pandas analysis and a chart out.

The model installs what it needs and writes files; you get the artifact back.

SCRAPE AND ANALYZE THE WEB SAFELY

A browser box plus a parsing box, feeding clean data to a model.

Playwright and CDP are ready in the browser box; parsing runs in a separate one.

GIVE EACH USER A PERSISTENT SHELL

A live, stateful shell per user, branch, or agent.

State survives stop and restart, so every user or branch keeps its own environment.

don't want to run it?

The same runtime, hosted.

BoxLite Cloud is the same SDK and the same core API, reached with a URL and an API key — managed storage, managed box lifecycle, nothing for you to operate.

BoxLite Cloud

New accounts start with $100 in credits. Plans from $19/month with included usage, or run wallet-only with no subscription.

Explore Cloud

faq

The questions people actually ask.

Those are cloud-only. BoxLite starts as a library that runs on your own infrastructure, including your laptop — no account to start, no data leaving your machine. If you later want it hosted, BoxLite Cloud runs the same runtime behind a URL and an API key; the choice stays yours.