Mitos cookbook

Isolated, forkable computers for your AI agents.
Four recipes to get from zero to a forked swarm.

01Parallel attempts with fork()

Do the slow setup once, then fork the live sandbox into N warm siblings. Each daughter shares the parent's memory pages until it writes, so every attempt starts from the same ready state in milliseconds.

import mitos

sb = mitos.create("python")
sb.exec("pip install -q requests")      # shared setup, paid once

attempts = sb.fork(4)                      # 4 warm copies of this exact state
for i, fork in enumerate(attempts):
    fork.exec(f"python solve.py --strategy {i} > /workspace/out.txt")

# read each fork's /workspace/out.txt, keep the winner, terminate the rest
sb.terminate()
Quickstart →

02A sandbox tool for smolagents

Give a smolagents agent a tool that runs code inside its own microVM instead of your machine. Same pattern works for LangGraph or a hand-rolled harness.

from smolagents import CodeAgent, InferenceClientModel, tool
import mitos

sb = mitos.create("python")

@tool
def run_python(code: str) -> str:
    """Run Python code in an isolated microVM sandbox.

    Args:
        code: The Python source to execute.
    """
    sb.exec(f"cat > /workspace/cell.py <<'EOF'\n{code}\nEOF")
    return sb.exec("python /workspace/cell.py").stdout

agent = CodeAgent(tools=[run_python], model=InferenceClientModel())
agent.run("Compute the first 10 Fibonacci numbers")
More agent patterns →

03Migrating from E2B

The create/exec surface maps one to one. What you gain is the part E2B does not have: forking a running sandbox.

# before: e2b
from e2b_code_interpreter import Sandbox
sbx = Sandbox()
sbx.run_code("print('hello')")

# after: mitos
import mitos
sb = mitos.create("python")
sb.exec("python -c \"print('hello')\"")
a, b = sb.fork(2)                          # the part you could not do before
Full migration guide →

04Self-host on Kubernetes

Everything above runs unchanged on your own cluster. Any Kubernetes with KVM nodes works; bare metal is a first-class target. Your data never leaves your infrastructure.

# point the SDK at your cluster instead of the hosted endpoint
export MITOS_BASE_URL=https://mitos.internal.example.com
Install guide →

Give your agents their own computers

Get an API key at mitos.run $5 credit to start. Open source, Apache-2.0, self-hostable.