How dots work: cloud computer, browser and plugins

A dot runs on GPT-6 Astra and operates inside its own environment: a cloud computer with a browser, storage and the ability to execute code. You supply objectives and permissions; the dot gets its tools from OpenAI’s plugin ecosystem and does the work without occupying your machine.

The five moving parts

Understanding these five explains almost everything a dot does.

1. The model: GPT-6 Astra

The reasoning comes from GPT-6 Astra, OpenAI’s frontier model. The model decides what to do; that is not the same thing as the agent. If you want to track model behaviour changes, the GPT-6 Astra system card publishes the change log.

2. The cloud computer

Every dot has its own computer. That is where the work lives: files, programs, runs. The point is isolation — your machine and its contents stay separate unless you decide to connect them.

You can open the dot’s computer at any time to inspect what it did. That is what makes the work auditable.

3. The browser and credentials

The dot browses the web on its own. To sign in to supported sites, it can use saved passwords without exposing them to the model, according to OpenAI. That detail matters: the credential never enters the model’s context.

4. The plugin ecosystem

dots connect to more than 4,000 apps through plugins. You choose which apps they can access and manage those permissions from ChatGPT’s Plugins tab.

One surprising detail: permissions are shared across dots, ChatGPT, ChatGPT Work and Codex. They are not exclusive to the dot. If you already had broad connections in ChatGPT, your dot starts with them.

On top of that, a dot can have its own Slack identity, with a separate account and its own access in those services. That makes it a distinguishable actor rather than something always acting with your credentials.

What it cannot have at launch is a standalone email address: it can use your connected personal account, but not an independent one.

5. The decision rules

A dot does not behave the same way in every context. It ships with built-in rules defining when it can move ahead on its own and when it has to ask for approval, and you can add custom rules to allow, require approval for, or block specific actions. Built-in safety requirements always apply on top of your rules.

Your computer: optional and off by default

The dot works on its own cloud computer. Access to your local computer is optional and starts turned off.

To connect it you use the ChatGPT desktop app on that machine and confirm “Allow access”. From then on the dot can reach files and work on that computer from any of its messaging channels. To cut it off, confirm “Revoke access”.

With local access enabled, the dot can also:

• create Work or Codex tasks,
• use local skills,
• use your local browser when its cloud browser is blocked.

It can also create tasks in Codex cloud environments, with one condition: the environment has to be created in Codex first before asking the dot to use it.

Proactive research: the background mode

When you are not actively working with your dot, it enters what OpenAI calls proactive research. In that mode it uses your connected apps, but with tools restricted to read-only operations.

The practical consequences are clear:

• It cannot send messages on your behalf.
• It cannot modify the content of your apps.
• It cannot control your browser or your computer.

It is a conservative design choice: the dot can look and propose, but touching anything requires returning to active mode with your permissions.

Auto-review: the check before the action

Before running an action that could affect your accounts or share information, auto-review checks that the action complies with your instructions, your custom rules and safety requirements. The outcome falls into three buckets: it can continue, it needs approval, or you have to do it yourself.

Certain sensitive tasks — changing a password is the example OpenAI gives — always stay with a human.

Parallel work

A dot can take on a whole project and carry it forward while working on several others. There is no need to open separate conversations or direct every step: you can keep adding tasks and ideas as it goes.

The full loop, in order

1. You create the dot, name it and customize it.
2. You connect the apps it may use.
3. You give it an objective or an area of responsibility.
4. The dot works on its computer — in active mode when you are around, in proactive research when you are not.
5. Before sensitive actions, auto-review decides whether it proceeds, asks for approval, or leaves it to you.
6. You review the result in the activity view or by opening its computer, and give new direction.