First run

Everything below is optional. MangoStudio works if you skip all of it — the setup flow exists so that nobody has to guess what to configure first, not to gate the application behind a form.

From an installed binary to a browser

mangostudio setup

One command takes a fresh install to a working hub:

1. generates and stores a BETTER_AUTH_SECRET if none is configured;
2. asks whether to install the background service (see below);
3. starts the hub, or reuses one that is already running;
4. opens it in a browser and prints the address.

A hub that is already serving is never restarted. A state file left behind by a crashed one is cleared and replaced.

On a machine with no display — over SSH, in a container, on a headless server — nothing pretends to open a browser. The address is printed along with a port-forward line:

MangoStudio is running at http://localhost:3001
No browser can be opened from here. Open that address on this machine, or
forward the port: ssh -L 3001:localhost:3001 <this-host>

In a script, --service or --no-service is required. Installing a service unit is a change to the machine that outlives the command, so with nobody at the keyboard the answer is refused rather than assumed in either direction:

mangostudio setup --no-service --no-open

--no-open prints the URL instead of opening anything.

The setup flow in the browser

The first time you sign in, MangoStudio opens Setup instead of the chat page. It has six steps, all of which can be skipped, and none of which is the only way to do what it does.

Progress belongs to your account, not to the browser: closing the tab and coming back resumes where you were, on any device.

What “resume” actually means

The flow does not remember which step you were on. It reads what the machine says now and opens at the first question still unanswered. So:

• delete the folder you chose, and it reopens at Folder;
• sign an agent CLI out, and it reopens at Agents;
• install Node from somewhere else, and Toolchain is simply done.

A request that has not answered yet reads as Checking, never as not done — a slow network cannot drag you backwards through steps you finished.

Skipping

Skip for now marks one step as skipped and moves on. Skip setup records the whole flow as finished and takes you to the page you were originally heading for. Neither sends a message or creates a chat.

To go through it again: Settings → General → Run setup again. That clears the progress record and nothing else — your chats, your vendor sign-ins and your machine settings stay exactly as they are.

Steps that cannot be done from where you are

Two steps depend on where the browser is, not on the account:

• Always on installs a service unit for the hub’s own process. If you opened MangoStudio from another computer — or the platform has no per-user service manager — the step says so and gives you the command to run there instead.
• Agents offers the agent CLIs installed on the machine you chose. Those sign in with the vendor’s own account, in a terminal on that machine; MangoStudio never sees those credentials and never asks for them.

A second person on the same hub

Agent CLI sign-ins on the Local machine belong to the operating-system account the hub runs as. Once a second MangoStudio account exists on the same hub, those sign-ins stop being offered — they are not the second person’s to use, and MangoStudio will not pretend they are by sharing one vendor account between two users.

The second person’s options are to use a model provider key, or to add a machine of their own from Environments, which gives them their own isolated agent sign-ins. The setup flow says this in the Agents step rather than showing a runner that every send would refuse.

The first chat

The last step creates one chat in the folder you chose, pointed at the runner you chose, and sends the message in the box. It is always an explicit click — nothing is sent by arriving at the step.

Two things it deliberately will not do:

• Create a second chat. The chat reference is saved the moment creation returns, before anything is sent. Reloading mid-flight reopens that chat.
• Send the same prompt twice. Before sending, it re-reads the transcript; a prompt that is already there means the send was accepted and only the answer was lost, so it waits rather than asking again.

A chat existing is not evidence that anything answered. The step is only done once the transcript actually carries a reply — or once you skip it.

If the runner is an agent CLI you have not used before, the third-party notice appears here, and the workspace-trust question appears the first time a vendor is pointed at a folder. Both are the same dialogs the composer raises, answered once.