July 2026
Dispatch an autonomous agent to a VM with one command. It replicates itself there (the binary scp's itself), runs detached, and keeps working while you walk away — then you attach back any time to inspect and steer it. One static binary, no Python on the target.
The worker runs a genuine tool-use loop — read files, write files, run shell — until the goal is met. Works with Claude (Anthropic) or any OpenAI-compatible endpoint such as OpenRouter, so you choose the model. The whole loop is written in machin (MFL), no SDK.
Hard token and iteration budgets freeze a runaway job. Shell is off by default and file access is confined to the job's own working directory. Every model turn, tool call, result, and token cost is journaled — the audit trail you attach to or pull home.
Turn on the confirm-gate and the agent parks on any destructive command (rm, dd, git push, DROP TABLE…) awaiting your approval — approve, deny, or stop, out-of-band. A deny budget auto-halts an agent that keeps trying variants. Since roam is non-interactive, a human and a supervising agent drive the exact same approval interface.
An optional goal-verify pass sends every completion to an independent judge that decides from evidence — the real working directory and the action log — not the agent's self-report. A failed check sends the agent back to work; it's fail-open, so a shaky verifier never blocks correct work.
Every command is one-shot: JSON on stdout, semantic exit codes, no prompts. Status and a live journal stream come back over plain ssh. A person on a laptop and an orchestrating agent use the identical surface — dispatch, watch, steer, approve, stop.