ChatGPT Dots Setup Guide: What to Connect, What to Lock Down, and When to Skip

TL;DR: Should you set up a dot?

Set up a ChatGPT dot if your plan already includes one and you give it a low-stakes recurring job, like a weekly summary or inbox triage drafts. Do not upgrade from Plus just to get one. If the job is the same steps every time, a scheduled task or Zapier is enough.

OpenAI describes a dot as “an always-on agent in ChatGPT that can take on ongoing responsibility and keep making progress between conversations.” It runs on GPT-6 Astra, has its own cloud computer, works across the apps you connect, and remembers context (OpenAI Help Center).

The premise: “keeps working between conversations, with memory” makes a dot useful and risky in equal measure, so the choices below matter.

Facts are current as of October 19, 2026 (dots launched September 29). One honest line: we have not used a dot. This guide is built on OpenAI's documentation and named early reports from WIRED and Zapier. Anything labeled “our judgment” is exactly that, an opinion and not a test result.

What a dot actually is (and what it is not)

A dot owns an ongoing responsibility (“keep track of my support queue”) and picks its own next steps as things change. Per OpenAI, it can also start ChatGPT Work (delegated tasks) and Codex (software tasks) itself. Those count toward Work and Codex limits.

Two details. First, the cloud computer runs Linux with a real Chrome browser, per Harry Guinness at Zapier, who used a dot on a Pro account for a few days. Zapier is an OpenAI integration partner and an automation vendor we also discuss, so read his account with that in mind. Second, per OpenAI's privacy FAQ, a dot can research in the background and save private notes, but those research tools cannot send messages, change content through plugins, or control a browser or computer.

Who can get a dot, and what it costs

Eligibility comes from OpenAI's Help Center and Learn docs. The dollar figures come from Zapier (October 13, 2026) and are echoed by WIRED and Lindy. We could not confirm them on an OpenAI pricing page, so read them as “Zapier lists.” Business workspaces start at two seats (OpenAI pricing page).

OpenAI says your first dot is “included in your Pro or Business Premium plan at no extra cost.” Chatting with the dot does not count toward ChatGPT usage limits. Allowance is free or extended for the first month (OpenAI's pages word that differently). Terms after that are unannounced, and more dots and capacity are promised with no price.

Rollout “may take several days,” so a missing Create your dot button at chatgpt.com/dots is not your fault. Start on desktop: OpenAI's pages disagree on creating one from mobile, and mobile web is not supported.

Five Business Premium seats at Zapier's rates come to $625 a month, or $500 billed annually (our arithmetic).

Setup in order: the first 30 minutes

Steps follow OpenAI's getting-started article and privacy FAQ. Menu labels differ between OpenAI's docs.

1. Turn on multi-factor authentication. OpenAI recommends it, and the dot inherits everything your account can reach.
2. Check existing plugin connections first. Per the privacy FAQ, permissions are shared across dots, ChatGPT, Work and Codex, and existing connections are reused.
3. Check “Improve the model for everyone” in your data settings (why: see the “can't undo” section).
4. Create the dot at chatgpt.com/dots or in the desktop app. Name it; the handle starts as @yourname-dot.
5. Connect apps via Customize > Plugins, using the list below.
6. Set Custom rules. They are optional. Set them anyway.
7. Leave local computer access off. It starts off. Allow access and Revoke access live in the desktop app, for one personal computer at a time.
8. Give it one job, written with four parts: what to check, when, what deserves a message, where results go. OpenAI's sample works as a template: “draft suggested replies for my review, but don't send them.” For scheduled jobs, include the time zone and an end date, and ask the dot to confirm the schedule.

Pause is in the ••• menu. Reset deletes the dot, including conversations, saved memories and scheduled tasks.

Which apps to connect first and which to hold back

Sort apps by what a wrong action costs, not by how useful the app sounds. This table is our judgment, built from OpenAI's descriptions and Guinness's report.

Four facts sit behind that table.

• Read access may not include change or send, per OpenAI. Check what each plugin grants.
• Connected apps are reviewed proactively. Per the Help Center, the dot can form memories from them “even when you haven't asked.”
• Email is a skeleton key. Guinness reported his dot signed into accounts using email codes without asking, because it had email access. See what access should an AI agent have.
• Slack is narrower than it sounds. Per OpenAI's Learn docs, adding the dot to a channel does not start monitoring, and by default it responds only to you.

The four Custom Rules, and which to use for what

Custom Rules set when a dot may act without checking with you:

The last row is partly forced. Per OpenAI, password changes and transfers between financial accounts are handed back, and permanent deletion and new security-sensitive access always need confirmation. On payments, one line: saved-card purchases need your approval; AI agent spending limits covers the rest.

OpenAI's reality checks matter more than the table does. Custom rules cannot override built-in safeguards or Auto-review (a check on certain planned actions, such as the recipient and message before an email). Asking for a draft does not authorize sending. Approving one message does not grant ongoing permission to contact people. And “your dot can make mistakes, including when following your rules.”

Guinness's report shows why: his dot sent an email without asking, but only after he had approved a previous send on the same thread. Approval can generalize, which is the whole problem. Keep sends on Ask before. Theory lives in AI agent permission to act without asking.

What you can't undo: memory, training data, and disconnecting

• Disconnecting an app does not delete what the dot already got. Per OpenAI, it “stops new information sharing through that connection. But information a dot has already learned remains in its own context.” To remove it, delete the dot.
• You cannot view, edit or delete individual memories. Only the whole dot. Turning Memory off stops sharing with ChatGPT but does not delete what the dot already received (privacy FAQ).
• Pause is narrow. OpenAI: “Pause stops your dot's current main task. It doesn't stop every delegated task or cancel future scheduled runs.” Stop delegated tasks in Activity; disable recurring ones in Scheduled.
• Deleting does not undo actions. It “doesn't undo changes already made in connected apps or recall messages already delivered to other people” (OpenAI Learn docs).
• Personal plans have a training default. On personal plans, the “Improve the model for everyone” setting controls whether your dot's conversations, actions and connected-app data may be used for training. Business, Enterprise and Edu are not used by default. Pro is a personal plan, so check it before you connect anything. OpenAI says human review can occur in limited circumstances even with the setting off, and that it does not train directly on background research notes.
• Prompt injection is reduced, not removed. OpenAI says its protections “do not eliminate it.”

When to skip a dot (and when a plain automation is enough)

Four questions:

1. Are the steps identical every time? Use a ChatGPT scheduled task, Zapier or n8n. A dot earns its keep on judgment over changing context.
2. Does a wrong action cost real money or a customer relationship? Keep a human in the loop or do not connect that app.
3. Is your plan and region eligible? Pro in the EEA, UK or Switzerland cannot get one.
4. Do you need predictable per-task cost or an audit trail you control? Post-month-one terms are unpublished. Wait.

Skip it if trying it means connecting email-with-send and customer data, if you only want a fixed routine, or if you are on Plus or below. Our judgment: do not upgrade solely for a dot until OpenAI publishes post-launch limits.

The reliability evidence is thin and mixed. Guinness reported his dot flagged routine payment and delivery notices but missed an urgent judgment email from his editor. His verdict: “it's still not ready to be the only tool that runs your calendar or inbox.” WIRED's Reece Rogers, a consumer-use reviewer, wrote after a few days that “this nascent feature is still rough in some areas,” though he found one research task's results “solid.” Two reporters is not a benchmark.

If a dot needs deterministic steps, Zapier's plugin can bridge that gap. For the decision itself: how to tell if a task needs AI or just automation.

A safe first week

This is your trial protocol, not our results.

• One job, two apps, read-heavy connections.
• Outbound actions on Ask before: drafts only.
• Review the Activity tab after every run, and check Scheduled for what is queued.
• Read the output. “Completed” does not prove the result was delivered or achieved, and OpenAI advises fixing missing details before expanding the dot's responsibility.
• Note daily what it missed and how much allowance it used.
• Set a success test first, like “saves an hour or removes a weekly chore with no cleanup.” If it fails, Pause. If it fails twice, Reset.

The AgentsExplained newsletter sends short, plain breakdowns like this each week.

FAQ

What are ChatGPT dots?

A dot is an always-on agent in ChatGPT, powered by GPT-6 Astra, with its own cloud computer. It works across apps you connect and remembers context.

Is a ChatGPT dot available on Plus or Free?

No, not as of October 19, 2026. Dots are included with Pro (outside the EEA, Switzerland and the UK) and rolling out to Business Premium, with an off-by-default beta for Enterprise, Edu and Healthcare. Free, Go, Plus and standard Business do not have them.

How much do ChatGPT dots cost?

Your first dot is included in Pro or Business Premium at no extra cost, per OpenAI. Zapier lists Pro at $100, $200 and $500 a month and Business Premium at $125 per user per month ($100 annually). Allowance terms after the first month are unannounced.

Can a dot access my computer?

Only if you allow it. Local computer access starts off. You can enable it in the desktop app and revoke it later, for one personal computer at a time, which must be online with the app open.

Does disconnecting an app delete what my dot learned?

No. Per OpenAI, disconnecting stops new information sharing, but what the dot already obtained stays in its context. Individual memories cannot be edited or deleted, so the only fix is deleting the dot.

Does OpenAI train on my dot's data?

On personal plans, it depends on your “Improve the model for everyone” setting, which controls whether dot conversations, actions and connected-app data may be used for training. Business, Enterprise and Edu are not used by default. Human review can occur in limited circumstances either way.

Can a dot spend money or change passwords on its own?

No. Per OpenAI, built-in safeguards hand back password changes and money transfers, and saved-card purchases need your approval. Custom Rules cannot override that.

Can I use a dot in Slack?

Yes. Per OpenAI, a dot can work in Slack and can have its own Slack identity. By default it responds only to you, and adding it to a channel does not start monitoring.

Is a dot different from ChatGPT Work or Codex?

Yes. Work handles a task you delegate now or on a schedule, and Codex builds software. A dot owns an ongoing responsibility and follows up between conversations. It can start Work and Codex tasks itself.

Bottom line

• Eligible, with a low-risk recurring job: try one. Read-only apps first, Ask before on anything outbound, review every run for a week.
• On Plus or below: wait for published post-launch limits.
• Fixed routine: automate it with a scheduled task, Zapier or n8n and skip the dot.

This is a decision guide built on documentation, not a review. For the wider field, start with how to build an AI agent without coding.