Docs

Configure agents

config.json on the office computer says which agents mobd runs and where. Everything else about them is set in Mob and shared by every device.

Install the agents

Install each agent you want on the office computer, as the user that runs mobd, and set it up once in a terminal: sign in with its account, or configure an API key the way the agent supports it.

Agent	Install	In `runners`
Codex	`npm install -g @openai/codex`, then `codex` to sign in or use an API key	`codex`
Claude Code	`curl -fsSL https://claude.ai/install.sh \| bash`, then `claude` to sign in or use an API key	`claude`
Pi	`npm install -g --ignore-scripts @earendil-works/[email protected]`, then set up its model provider	`pi`

mobd checks that each agent is installed in a supported version; Mob's machine settings list every agent with its state and, when one is missing, the command to install it. The agent's own sign-in or API key is not checked: a problem there shows as the agent's error on the first turn.

config.json

mobd init     # creates ~/.mobd/config.json; never overwrites it

Open ~/.mobd/config.json in an editor (open -e on macOS, nano on Linux):

{
  "listen": "0.0.0.0:8080",
  "workdir": "/Users/you/Projects",
  "runners": ["codex", "claude", "pi"]
}

runners: the agents to enable. mobd finds codex, claude and pi on your PATH; otherwise set codexBin, claudeBin or piBin to the full path (command -v codex shows it).
workdir: the folder the agents work in, as a full path: /Users/you/… on macOS, /home/you/… on Linux. JSON does not expand ~ or $HOME.
listen: 0.0.0.0:8080 lets devices on your network and VPN connect. With Tailscale or a Cloudflare Tunnel only, use 127.0.0.1:8080.
Run mobd restart after a change.

Every setting is in the configuration reference.

Agent settings in Mob

Once a device is paired, open the machine in Mob (Settings → Machines, then its Agents settings) to set, for every device at once:

the default agent for new conversations;
the models offered for each agent, and its default model and effort;
permissions: Run without asking (the default), or Ask before commands and edits, where the agent waits for your approval in Mob.

Changes apply from the next turn, without a restart. A conversation keeps its agent; its model and effort can be changed in the message box.

Linux desktops

On Linux these settings shape the conversations' desktops (all optional):

Setting	Default	What it does
`maxRunning`	`4`	Turns running at once; others wait
`desktopSize`	`1920x1080`	Size of every desktop, in points
`desktopScale`	`1`	`2` for HiDPI
`desktopSleepAfter`	`10m`	Idle time before a desktop sleeps; `0` never
`desktopKeyboardLayout`	the computer's	Keyboard layout of every desktop
`maxDesktops`	`0` (no limit)	Desktops open at once; new conversations wait
`memoryWarnPercent` / `memoryFloorPercent`	`15` / `5`	Available memory to warn at / below which no new desktop starts

How the desktops work: Desktops & live view.

Security

Open it on trusted networks only. Every device must pair, but plain HTTP is not encrypted. Never forward port 8080 to the internet; across networks see Remote access.

The agents run as the user that runs mobd, with that user's files and logins. With Run without asking they act without confirmation; choose Ask before commands and edits for an agent you want to supervise.