Skip to main content
A machine is a Linux computer that belongs to you. An agent runs on it, writes to its disk, and finds what it wrote there the next time — in the next turn, in the next chat, tomorrow. One is created for you when you sign up. Open Machines in the sidebar and pick Your machines to see it.
The Machines page showing one running machine

A machine: what it runs, what it has, what it has used, when it was last active

What it is

The agent’s home is /home/daytona. Inside it: The card above reads 2 vCPU · 4 GB RAM and 13.8 MB of 30 GB used. The same numbers come back from the app’s own API, alongside the harness the machine runs:
harness_ref is the interesting field: a machine runs one harness, at one ref. Nothing else is checked out beside it, so there is nothing for ~/harness to collide with and no path to disambiguate. A second harness of yours runs on a machine of its own.

Start and stop

The machine detail page with the power section

A machine's power section, from the machine page

The app says it in one line: a stopped machine keeps its files and comes back where it left off. Stopping is not deleting. What stops is the running state — the processes, the memory — and what stays is the disk. On a plan that does not keep machines running, the machine also stops itself once nothing has used it for a while; auto_stop_minutes in the live half above is that timer — 30 on the free plan. Starting it again is one button, or simply sending a message.
Anything the agent left running dies with the machine, and only with it. A program started in one turn keeps running after that turn ends, through the next chat and into tomorrow — until somebody presses Stop, or the plan stops keeping machines running. Nothing reaps it on a schedule. Write a long-running program so that starting it twice is harmless and starting it after a stop picks up where it left off, and “has it stopped?” stops being a question anybody has to answer.

What survives, and what does not

The projection of a published version leaves a dirty checkout alone: it will not overwrite work sitting in the sandbox with the published commit. That is what makes editing a tool or a prompt in place a reasonable way to develop one.

One machine, or several

A machine runs one harness. A harness can be given more than one machine — on the Your harnesses tab, the card says “Give it another computer to run on. The ones it already has stay.” — and each of those machines is a separate computer with a separate disk.
The Your harnesses tab showing one harness on one machine

A harness, and how many machines it is on

How many you may have, and whether they are kept running, comes from the plan:
On the free plan a machine still keeps its disk across a stop. What it does not do is stay up while you are not talking to it, which is the thing a long-running program needs.

Which machine a chat runs on

The picker at the top left of a chat is the machine picker, not an agent picker. It reads the machine’s name, and under it the harness and ref that machine runs.
The machine picker open in a Splox chat

The machine picker in a chat: the name, then the harness at its ref

Switching machines switches computers: a different disk, a different checkout, possibly a different harness. A chat stays on the machine it was started on.

Harness

What the machine runs, and how a version reaches it.

Memory and what persists

The four places a fact can live between runs.