Skip to main content
Your agent works on a real Linux machine, and you can go and look at it. Two doors, both in the chat you are already in: an editor in the browser, and an SSH command you paste into a terminal. Both are for looking. Reading what the agent wrote, checking that a file has what you think it has, watching a program’s log, fixing one line by hand. You do not have to use either of them to change your agent — you ask it, and it edits itself. These are for the times you want to see for yourself.

​
Where they are

Open the panel on the right of a chat — the icon in the top-right corner — and find the Sandbox section.
A Splox chat with the inspector panel open on the right

The chat inspector, open on a finished run

The Sandbox card showing Open in VS Code, the SSH command and Revoke access

The Sandbox section: an editor, an SSH command, and one button that closes both

Three controls, and one sentence under them that is worth taking literally: Revoking access ends both the SSH command and the editor session. If the section says The sandbox appears once the agent runs code, this chat has never needed a machine yet. Ask it to do something and the doors appear. If it says This sandbox is asleep. Opening it wakes it up, the machine is stopped. Either button starts it, which takes a minute or so — the button reads Waking the sandbox… while it waits, and gives up after 90 seconds with The sandbox did not wake up in time, which means come back and press it again. Starting a machine this way counts against your plan exactly like pressing Start on the Machines screen does.

​
Open in VS Code

The button opens a new browser tab with a full VS Code — the real editor, running on your machine — on /home/daytona, the agent’s home directory. Not on a project subfolder: the whole home, so harness/ is there next to everything the agent has written.
code-server open on /home/daytona with harness/programs/splox/prompts/assistant.md

VS Code on the machine, with the agent's own prompt file open

The first time it opens, VS Code asks whether you trust the folder. It is your machine and your agent’s files; say yes. What it is good for:
  • Reading what the agent wrote. harness/programs/splox/prompts/assistant.md is what your agent is told, in English. Two clicks from that tab.
  • Fixing one line. A word in a prompt, a number in a program. Save it and your next message in the chat reads the new file — nothing to restart. (It stays on this machine until it is published: see Versions.)
  • Looking at a file the agent produced without asking for it — a CSV, a report, a log — in an editor rather than in the conversation.
The link is the credential. There is no password on the editor: anyone who has that URL is inside your machine until you revoke it. Do not paste it into a ticket, a chat room or a screenshot.

​
SSH

The SSH button reveals a command and a countdown:
That random string is not a placeholder — it is the whole credential, carried as the username. Copy the line, paste it into your terminal, and you are on the machine as the user your agent runs as, in /home/daytona. It is good for the things a terminal is good for:
And for getting a file off the machine without going through the chat:
Three things to know:
  • The grant lasts an hour. The panel counts it down — 0h 59m remaining.
  • Only the newest command works. Pressing SSH again mints a new one and the previous command stops connecting, in the same second. If a command you copied earlier is refused, that is why.
  • You cannot publish from there. The checkout at ~/harness is readable and editable over SSH, but git push fails to authenticate: the token that lets a push through belongs to a run, not to your terminal.
    This is the seam between looking and authoring. Edit the file here if you like; ask your agent to publish it.

​
Closing both

Revoke access ends the SSH grant and the editor session together. Afterwards the command is refused —
— and the editor tab answers 502, because the process behind it is gone. The machine and everything on it are untouched; only the doors are shut. Nothing else closes them for you on your schedule. The SSH grant expires on its own after an hour, and the editor’s link is signed for an hour; the editor process itself shuts down once the tab has been idle for a while, and dies with the machine. If you opened a door on a machine somebody else can see the screen of, revoke it when you are done rather than waiting for any of that.

​
When a door does not open

​
The requests behind the buttons

For an agent reading this page, or anyone driving the app’s own API with a session cookie:
​
request
Does this chat have a machine, and what state is it in. {"exists": true, "state": "started"}. Reports only — it never starts anything.
​
request
One SSH grant, one hour. {"command": "ssh -p 2222 <token>@sandbox.splox.io", "expires_at": "…"}.
​
request
Starts the editor and returns a link to it: {"url": "https://8443-….splox.app"}. No expiry field — the link is signed on the runner’s clock, and this API will not invent a second deadline for it.
​
request
Closes both doors. 204, and 204 again for a chat that has no machine.
Both POSTs answer 202 {"state": "starting"} while a stopped machine comes up; the client is expected to ask again. Neither one ever creates a machine.

Files

What you will find when you get there.

Versions

An edit you made by hand, and how it becomes a version.