Safety and the sandbox
What it can never do, what it asks about, and how much disk it reaches.
An agent that runs commands needs limits that do not depend on the model behaving. motita has two layers, and only one of them is yours.
The floor
Some commands are refused every time: disk formatting and partition tools, shred, dd to an output, power-state commands, and rm, mv, cp or ln aimed at /, your home, a system path, or the folder that contains your workspace (including after sudo, xargs or find -exec). No key, variable or flag relaxes the floor.
Allow, ask, deny
- 1A commandThe agent wants to run it
- 2The floorFormatting,
shred,rm /…: refused every time - 3The verdictAllow, ask you, or deny
- 4The sandboxRuns confined, with limits
The floor never moves. The verdict and the sandbox are yours to tune.
| Verdict | When |
|---|---|
| Allow | The command changes nothing, or only the workspace: make, go build, npm test, your own ./scripts/*. It runs silently, because a policy that asks about those gets switched off. |
| Ask | It reaches the network or the system, writes outside the workspace, or cannot be classified. You are asked before it runs. |
| Deny | Strict mode and the line cannot be placed. |
The sandbox
Commands run confined, with no Docker. Memory, CPU time, processes, open files and output size are limited, optionally with no network inside the command. On Linux with Landlock, a command cannot write outside the directories it works in and cannot read motita's logins, configuration or your git credential store.
Disk access
Settings → Security chooses how much of the disk the agent reaches:
- Workspace only (default): file edits outside the session's folder are refused, and on Linux a command cannot read or list your files elsewhere (home,
/home,/mnt, ...). - Whole disk: both reach anywhere your account can. The page asks first and spells out the risk. motita's own logins stay hidden either way, and commands that work outside the workspace stop asking (one that names your keys still asks).
Where nothing confines a command (Windows, macOS, Linux without Landlock), the agent asks before the project's own code runs, unless you turn that off.
Each session also gets its own HOME, and cannot read the other projects and sessions of the same gateway.
The Sandbox section of the same page lists the services (a database, a web server) the agent runs in boxes of their own to test against, and stops them.
This computer's desktop
Off by default. Turned on in Settings → Security, after a confirmation that spells out the risk, the agent can look at the screen and use the mouse, the keyboard and the windows of the computer motita runs on: Linux (X11 or Wayland), macOS and Windows.
- Every step is asked about, a picture of the screen included, since it shows whatever you have open. Allowing every command for the session allows them all.
- Never from a background agent or a read-only session; while it is off the model is not even offered it.
- The pictures are saved under
.motita/previews/and shown with the result.