Infrastructure for agent swarms.

Mitos forks a running machine into a fleet of isolated microVMs, so one agent becomes a thousand.

Start with the open engine View on GitHub

Most sandbox platforms were built to run untrusted code safely. That problem is largely solved: microVMs are everywhere now. The pressure has moved to a harder one: running fleets of stateful agents in parallel for best-of-N attempts, RL rollouts, tree search, and large multi-agent evals.

Mitos is built for that. Instead of cold-booting or restoring a snapshot for every new environment, it forks a running Firecracker microVM: the live memory and processes of a warm machine are copied-on-write into hundreds of isolated daughters, each its own microVM, in about 27 ms and roughly 3 MiB apiece. One agent divides into a thousand.

The managed cloud is the product: we run the bare-metal fleet, keep pools warm, and hand you an API. The engine underneath is open source under Apache 2.0, the same code we run. That is your no-lock-in guarantee, not a side project: you can read it, run it yourself, and leave whenever you want.

Numbers you can reproduce

Every figure on this site comes from a script in the open-source repository. If a claim is not backed by code you can run, it is not here.

Open source

The engine, SDKs, and benchmarks live on GitHub under Apache 2.0. Issues, discussions, and pull requests are welcome. If you want to reach us, see the contact page.