One command carries the whole task
Name a harness, say what the work is, and the agent does it somewhere else:
npm install -g @munitoai/cli
munito agent dsh "add a retry to the HTTP client"Munito AI starts a machine, uploads your git worktree, runs the agent, applies the patch it wrote and deletes the machine. Nothing runs on your laptop, and nothing is left behind. In our own measurement the machine was ready in 772 milliseconds.
Eight harnesses, one key
Claude Code, Codex, OpenCode, DeepSeek Harness, Hermes, pi, OpenClaw and Prime Agent, each already pointed at our open models in the EU. One key pays for the machine and for the tokens.
Your own Anthropic or OpenAI key works too, and so does your own subscription: the sign-in happens inside the machine, over SSH, and never passes through our API. See Cloud agents.
Read the repository, then change it
An agent that can run commands can also read the code first. In one measured run, DeepSeek Harness answered a question about a repository in 2 seconds at 251 tokens a second on GLM-5.3-Flash, then wrote four files in 1 minute 54 seconds.
Open models are fast enough for the loop a coding agent lives in, and the loop runs beside the model rather than across the Atlantic.
Watch it work, or do not
The same machine carries a desktop, so you can open a screen and see the agent type. Leave the screen out and the run is a command that prints a patch, which is what a queue or a pipeline wants. The console lists every sandbox that runs an agent, and opens it.
A run ends with its machine by default. Keep it instead, and you have a persistent computer: a shell over SSH, a pause that holds the disk and the sign-ins, and a resume where the agent stopped.
Give it the smallest key you can
Every program inside the machine reads its environment, and so does the agent. Use a key you can revoke, with the permissions and the spend limit of this one task.
Cloud agents is a public preview, billed by the second on the same account as the inference API. See Cloud agents, the command line tool and what it costs, or talk to an engineer.