OpenAI dots Guide: Setup, Permissions, and Scheduled Work

OpenAI introduced dots on September 29, 2026 as always-on agents that can keep working toward an ongoing goal between conversations. A dot has its own cloud computer, can use information from apps you choose to connect, and can run scheduled work or background research.

The important part is not simply that a dot can continue working. It is deciding what it may read, what it may change, and when it must stop for your approval. This guide walks through initial setup, connected apps, scheduled tasks, Custom Rules, memory, local-computer access, and safer ways to test the feature.

OpenAI dots at a glance

Area	What to know
Availability	Dots are rolling out to eligible ChatGPT Pro and Business Premium users. Enterprise, Edu, and Healthcare workspaces need an administrator to enable the beta.
Where to create one	Use the ChatGPT desktop app or desktop web. You cannot currently create a dot on mobile web.
Work environment	A dot uses its own cloud computer by default. Access to your local computer is optional and starts turned off.
Connected apps	Plugin connections and permissions are shared across dots, ChatGPT, ChatGPT Work, and Codex.
Action control	Custom Rules can let the dot act, act only when pre-approved, ask each time, or hand the step back to you.
Deletion	Reset removes the dot’s conversations, saved context, and scheduled tasks, but separately created files or Codex threads may remain in their own locations.

How a dot differs from a normal ChatGPT conversation

A normal conversation is usually centered on the request in the current thread. A dot is designed to maintain responsibility for an ongoing project, notice new information, decide what should happen next, and return to you when a decision requires judgment. Its profile organizes work under In progress, Scheduled, and Completed, and you can open the dot’s computer to inspect or interact with its work.

Dots also support proactive research. A dot may review permitted connected information and create private notes even when you have not sent a new message. OpenAI says the tools available during proactive research are read-only: they cannot directly send messages, change app content, or control a browser or computer. Any follow-up action still has to pass the normal action rules and safety checks.

Conceptual view of connected apps, a private dot workspace, and a human approval checkpoint — Connected information enters a separate work environment, while sharing, purchasing, and changing data remain subject to rules and human review.

What you need before setup

An eligible plan and region: Pro rollout excludes the European Economic Area, Switzerland, and the UK at launch. Business Premium is available across supported ChatGPT regions. Access is gradual and may take several days to reach an eligible account.
A desktop session: Initial creation and messaging-channel setup require the desktop app or desktop web.
A narrow first responsibility: Define the source material, schedule, expected output, and actions that must remain under human control.
A data boundary: Do not paste passwords, payment details, or unnecessary customer data into chats or documents. Supported sign-in flows pause for you to enter credentials through a secure form.

OpenAI says the first dot is included with Pro or Business Premium at no extra cost. Conversations with a dot do not count toward regular ChatGPT usage limits. If the dot starts work in Codex or ChatGPT Work, those delegated tasks use the normal allowance for the relevant product.

How to set up OpenAI dots

1. Create the dot on desktop and give it one clear role

Start the onboarding flow in the ChatGPT desktop app or desktop web, then choose a name and profile. Keep the first responsibility specific. Instead of “manage all my work,” try: “Every Monday, review the connected project documents and prepare a priority draft, but do not message anyone or edit the source documents.”

Completion check: Setup is complete when the dot’s name and handle appear, you can open its profile, and the activity sections are available. If the feature is missing, confirm the plan, region, and gradual-rollout notice before changing subscriptions.

2. Connect one necessary app at a time

Connected apps define the information boundary for the agent. On mobile, OpenAI directs users to Customize → Plugins in the dot profile to review available connections and permissions. These plugin settings are shared with ChatGPT, ChatGPT Work, and Codex, so a connection may affect more than the dot alone.

Begin with one source that has an obvious read purpose, such as a calendar or project-document folder. Confirm what the dot can actually access before adding more services. Disconnecting an app stops future access but does not automatically remove information already incorporated into the dot’s context. Deleting that context requires resetting the dot.

3. Put recurring work in Scheduled

You can ask a dot to set a reminder or run a recurring check. Open the dot profile and use Scheduled to review active, paused, and completed tasks, then adjust the repeat schedule, time, or completion notifications.

A good scheduled instruction includes more than a time. Name the sources, output structure, and prohibited actions. For example: “At 8:00 a.m. on weekdays, review my calendar and the connected project list. Return three sections for deadlines, schedule conflicts, and decisions I need to make. Do not send external messages.”

4. Set an approval level for each action with Custom Rules

OpenAI’s help article lists four available behaviors:

Take action without asking
Take action if pre-approved
Ask before taking action
Hand off to you

For an initial trial, consider allowing independent reading and drafting while requiring confirmation or handoff for external sharing, sending, purchasing, deleting, installing software, or changing permissions. “Pre-approved” means that the specific action was explicitly requested in your prompt. It is not a permanent blanket authorization for similar future actions.

OpenAI Help Center documentation showing Custom Rules and approval options for dots — The Custom Rules choices in OpenAI’s official help documentation. Source: OpenAI Help Center, checked October 1, 2026.

5. Enable local-computer access only when the task needs it

A dot works on its own cloud computer by default. Local-computer access is optional and must be enabled from the ChatGPT desktop app on that computer. After you confirm Allow access, the dot can work with local files, start Work or Codex tasks, use local skills, and use a local browser when its cloud browser is blocked.

If the workflow can be completed through cloud apps, leave local access off. Enable it for a defined task, review the result, and use Revoke access when it is no longer needed. Camera, microphone, and screen access also require separate operating-system permissions, so grant only what the job requires.

Three practical workflows

1. Turn project changes into a weekly briefing

Inputs: Read access to project documents and a calendar, a Friday schedule, and a fixed briefing format. Ask the dot to identify newly changed decisions, deadlines for the following week, items without an owner, and schedule conflicts.

Flow: The dot reads connected sources and prepares a private draft. Keep Slack or email delivery disabled until a person has reviewed the summary. This limits the chance that a mistaken interpretation is distributed to the entire team.

Review: Check that claims link back to source documents, completed work has not been resurfaced, and owners or dates were not guessed.

2. Separate repetitive steps in a content workflow

Inputs: A folder for interview transcripts, editorial guidelines, prohibited wording, and a destination for drafts. When a new transcript arrives, the dot can prepare candidate clips, title ideas, and an episode description while publication and external sharing remain approval-gated.

Flow: The dot produces the repeatable artifacts, and the editor reviews facts and wording. Instruct it to cite transcript locations instead of presenting predictions such as “this will perform well” as evidence.

Review: Compare quotations with the transcript, remove personal information, and confirm that the final publish action still belongs to a person.

3. Investigate a small software issue and prepare a pull-request draft

Inputs: The target repository, reproduction steps, the files the agent may change, and required tests. If you want the dot to use a Codex cloud environment, create that environment before asking the dot to use it.

Flow: The dot can review the issue and logs, narrow the cause, delegate a code-and-test task, and bring back the result. Deployment, secret changes, package installation, and permission changes should require confirmation or direct takeover.

Review: Verify that the reproduction case is fixed, unrelated files are untouched, test evidence is present, and unresolved risks are separated from completed work. Do not automatically merge a pull request solely because the dot prepared it.

Permission, memory, and privacy details that matter

Disconnecting an app is not the same as deleting retained context. It stops new access but may leave information the dot already learned.
Individual dot memories cannot currently be viewed and deleted one by one. Reset is required to remove the dot’s own saved context.
Personal accounts should review the model-improvement setting. The “Improve the model for everyone” control determines whether eligible dot conversations and work may be used to improve OpenAI’s models. Proactive-research notes are not used directly for training, but information brought into an eligible conversation or task may be covered by the setting.
Business data follows different defaults. OpenAI says Business, Enterprise, and Edu workspace content is not used for model training by default.
Safeguards reduce risk but do not eliminate mistakes. Plugin permissions, Custom Rules, action review, and safety monitoring work together, but consequential outputs still need human review.

Troubleshooting checklist

No dots option: Confirm the plan and launch region, then check from desktop web or the desktop app. Gradual rollout means an eligible account may not receive access immediately.
The dot cannot read an app: Review the connection and scope under Customize → Plugins. Reconnect only the required service instead of broadly expanding permissions.
A scheduled result is wrong: Check the timezone, recurrence, data sources, and notification setting in Scheduled. Add an explicit output format to the task instruction.
The dot attempts an unwanted action: Stop the task and strengthen the relevant Custom Rule to Ask before taking action or Hand off to you. Check the connected service separately to see whether the completed action can be reversed.
You want retained context removed: Disconnecting the app is not enough. Read the Reset notice, reset the dot, and separately manage any files or conversations stored elsewhere.

Who should use dots?

Dots are most useful for people who repeatedly monitor the same information, coordinate a project across several apps, or follow a stable draft-review-approval workflow. A regular ChatGPT conversation or a narrower automation is a better fit for one-off questions, tasks with almost no connected context, or high-impact work that cannot be meaningfully reviewed by a person.

The safest starting point is small: one read-only source, one scheduled task, and confirmation for every external action. Expand access only after the outputs are consistently useful and easy to audit.

Official sources

Sources checked October 1, 2026. Availability, limits, and controls may change, so verify the current help article and the settings shown in your account.

Search This Blog

Stackfinch