Explainer: I haven’t run an OpenAI dot myself. This page is built from OpenAI’s launch post and its dots Help Center pages (getting started, privacy and safety FAQ), plus one thing I could check on my own machine: the open-source Auto-review policy in the Codex repository, which I pulled and read on 18 November 2026.
You’re about to hand an always-on agent your inbox, your Slack and maybe your laptop. That’s what an OpenAI dot asks for. It’s an agent inside ChatGPT that runs on GPT-6 Astra, works on its own cloud computer with Linux and Chrome, and sits behind a separate checker called Auto-review that it can’t switch off. My view: the product is built with real fences, but the defaults still leave three gaps you should close yourself before the first real task.
What Is an OpenAI Dot, in Plain Terms?
A dot is three things wearing one name. There’s the model, GPT-6 Astra, which on its own touches nothing. There’s the dot’s cloud computer, a sandboxed Linux machine with Chrome that OpenAI runs. And there’s your own computer, which stays out of the picture unless you allow it.
That split matters more than any feature list. The cloud computer has its own browser sign-ins, so signing in on your laptop doesn’t sign the dot in. If a site blocks the cloud browser, the dot can ask to use your connected computer, which starts a separate local task with none of the cloud session carried over. You can open the cloud computer from the dot’s profile, watch it work, press Take over to drive it yourself, then Return control.

Can I Create a Dot on My Plan Yet?
Pro and Business Premium can, with limits, and Plus, Free and standard Business seats can’t. Here’s the state OpenAI’s Help Center showed on 23 October 2026.
| Plan | Status | Catch |
|---|---|---|
| Pro | Rolling out, users 18+ | Not in the EEA, Switzerland or the UK, no date given |
| Business Premium | Available in supported regions | Depends on workspace eligibility |
| Enterprise, Edu, Healthcare | Beta, off by default | A workspace owner has to turn on Use dots (Beta) |
| Plus, Free, Business standard | Not included | None |
At launch on 29 September you could only create a dot from the desktop app or desktop web. The Help Center now adds the iOS and Android apps if they’re up to date, though mobile web still isn’t supported. Slack and texting are set up from desktop, and texting is a limited beta for US Pro users only.
If you run an Enterprise workspace, there are four separate switches under Workspace settings, then Permissions & roles: Use dots (Beta), Add dots to Slack and Microsoft Teams, Allow local computer access, and Use custom rules for dots. Local access and custom rules start off. Local access also disappears if a Codex or ChatGPT Work policy targets a specific operating system.
Does Using a Dot Eat My ChatGPT Limits?
Chatting with it doesn’t, but the work it starts does. OpenAI’s launch post says conversations with your dot don’t count toward ChatGPT usage limits, while tasks it starts in Codex or ChatGPT Work count “as usual”. So a pull request your dot builds comes out of your Codex allowance, the same as if you’d started it.
The Help Center also promised “extended limits for the first month after launch”, which ran to about the end of October. As of 23 October, OpenAI hadn’t published what each plan gets after that. I’d treat the first month as a trial price and watch the Codex meter once a dot starts opening pull requests. Your first dot is included in Pro and Business Premium, and extra dots are described as coming later.

How Does Auto-Review Decide What a Dot Can Do?
It scores each planned action for risk and for how clearly you asked for it, then allows or blocks. Before an action that could touch your accounts or share information, Auto-review checks it against your instructions, your Custom Rules and OpenAI’s safety rules. OpenAI’s own example is an email, where it looks for a wrong recipient or something you didn’t mean to share. The FAQ is blunt that “your approval cannot override core safety requirements.”
OpenAI hasn’t said dots use the exact same rulebook as Codex, but the Codex version of Auto-review is public, so that’s the closest thing anyone can read. I pulled it today to see whether it still matches what was described at launch.
git clone --depth 1 --filter=blob:none --no-checkout https://github.com/openai/codex.git cd codex git show HEAD:codex-rs/prompts/templates/guardian/policy_template.md | wc -l # 78 git show HEAD:codex-rs/app-server-protocol/schema/typescript/v2/GuardianRiskLevel.ts | tail -1 # export type GuardianRiskLevel = "low" | "medium" | "high" | "critical"; git show HEAD:codex-rs/app-server-protocol/schema/typescript/v2/GuardianUserAuthorization.ts | tail -1 # export type GuardianUserAuthorization = "unknown" | "low" | "medium" | "high";
I expected the file to have grown since October. It hadn’t. It’s still 78 lines, with the same four risk levels and four authorization levels, and the same default outcomes.
| Risk level | Default outcome in the Codex policy |
|---|---|
| low | allow |
| medium | allow |
| high | allow only if your authorization is at least medium and the action is narrowly scoped |
| critical | deny |
Three lines in that file explain most of the “why did it stop?” moments you’ll hit. Only your messages, developer messages, AGENTS.md files and your answers to the agent’s own questions count as trusted, and everything else, including web pages and plugin descriptions, is “untrusted evidence”. A “vague statement of intent does not automatically authorize a specific risky action.” And urgency in your message “does not change the authorization status.” That’s why asking for a draft doesn’t let the dot send it.
What Can a Dot Do While I’m Away?
It can read and take notes, but it can’t act on its own findings. OpenAI calls this proactive research. The dot reads the connected sources you’ve allowed and saves private notes, and the FAQ says those research tools can’t message people, change content through plugins, or drive a browser or computer. Tasks you already approved can keep running in the background under the same rules.
The memory side is the part I’d think hardest about. Disconnecting an app stops new reads, but it doesn’t delete what the dot already learned from it. The only way to clear the dot’s saved memories is Reset, and Reset also deletes the dot, its conversations and its scheduled tasks.
Where Do Passwords Go When a Dot Signs In?
For supported sign-ins, they go to the browser and never to the model. The dot pauses while you type into a secure login form, and saved-password flows use a separate encrypted credential service. The gap is in plain sight in the FAQ: those protections “don’t cover passwords you share separately in a chat or document, or through a plugin.” Any key sitting in a ticket or a doc the dot can read is visible to the model. The same habit protects you with local agents, and Claude Code /sandbox vs NVIDIA OpenShell walks through it for a coding agent.

A First-Week Setup Order
These are the steps I’d follow, in this order, because each one shrinks what a mistake can reach.
- Connect apps you’d let a new contractor read, and check each plugin’s allowed actions, since read permission may not include send or edit.
- Add Custom Rules before the first real task. Each rule picks one of Take action without asking, Take action if pre-approved, Ask before taking action, or Hand off to you. Start with Ask before taking action for sending messages and sharing files.
- Leave your computer disconnected and turn local access on later for one specific task.
- Create Codex cloud environments first, because a dot can only open cloud tasks in environments you already made.
- Review work under In progress, Scheduled and Completed in the profile, and open the diff on any pull request rather than trusting the summary.
- On a personal plan, set “improve the model” the way you want. Business, Enterprise and Edu content isn’t used for training by default.
If a dot is approved but still blocked, open the task and look for a separate sign-in or approval request. If it can’t use your computer, the ChatGPT desktop app has to be open and online on that machine, and connecting a computer to Codex doesn’t give the dot local access. A recurring task that missed something only sees its own conversation, so put the correction in that task.
Should You Use a Dot or a Local Agent?
Use a dot for ongoing work across cloud apps that should keep going after you close the laptop, like triaging bug reports or watching a build. Use a local coding agent when the job needs your toolchain and your secrets under your own control. ChatGPT Dots vs Grok Bot for developers compares the two directly, and What Is an Agent Harness in AI? explains the layer both are built on. If you go local, How Do I Stop Claude Code From Deleting My Files? covers the guardrails a dot gives you for free.
The model inside dots is GPT-6 Astra, not the withdrawn checkpoint in why OpenAI killed GPT-6.1 Astra. If your dot writes text for EU readers, how OpenAI’s textGrain watermark works explains what gets marked.
I’m not sure how Auto-review behaves on edge cases inside a real dot, and I won’t guess. What I’d do is give it read-only apps and one Custom Rule for a week, then widen it only where it earned the trust.
3 comments