Mitos

Mitos is a service for forking a running microVM into an agent swarm. It is described as providing isolated, forkable computers for agents, and the site frames it around workloads that need many parallel isolated environments.

Its core function is a live fork: a warm virtual machine can be copied along with its memory and processes, not just its disk. Each fork is its own microVM, and the page says the forks share the parent’s pages until they are modified, with about 3 MiB per subagent after forking and about 27 ms per fork. The examples and use cases named include best-of-N attempts, reinforcement learning rollouts and fan-out, tree search, multi-agent evaluations, and running model-written or otherwise untrusted code in a microVM with its own kernel. The page also says a warm machine can become a swarm and later collapse back to zero.

Mitos is accessible through Python and TypeScript SDKs, a CLI, and an MCP server. The examples show creating a warm sandbox, forking it into a swarm, and executing commands in the resulting microVMs. The page also says the managed service is fully managed: the company runs the cluster, while the user calls an API. It states that there is no need to manage clusters, KVM hosts, or snapshot storage, and that pools stay warm so forks land in milliseconds.

The product is open source under Apache 2.0, and the page says the same engine runs in the managed service and can be run independently. Pricing is described as prepaid credit with per-second metering, along with hard spend caps. The page also says the engine is open source, there is no lock-in, and the managed offering is backed by a warm bare-metal fleet.