GPT DotFIELD GUIDE
Home/Guides
Guides

How to connect Gmail, Slack, Drive or GitHub to your ChatGPT Dot: permission levels, a first test and fixes for five common failures

Connect Gmail, Calendar, Slack, Drive, Notion, GitHub or Teams to a ChatGPT Dot: pick the job first, set each app's permission level (Always ask, Allow read actions, Allow low-risk actions or Allow all actions), put sends behind Ask before taking action, then run a one-minute write test. Plus fixes for a false "user cancelled", approvals that miss the write, watches that go quiet and the wrong account.

GPT Dot · Updated

The short version: how to connect apps to a ChatGPT Dot, what to set before it writes, and what to check when an app fails

In one line

Connecting apps works well for reading and summarising: a Dot can search Gmail, read your calendar, Drive, Notion and GitHub, and sum up your Slack mentions within a few minutes. Writing and watching work only partly. Sends, uploads and triggers fail separately from reads, so a lookup that works says little about a send. It suits people who start with one job and one or two apps, set the permission level for each app, and test the write path once before relying on it. It doesn't suit anyone who needs a new-mail trigger to follow a chosen Gmail account, a hosted IMAP mailbox, or a Dot that only reads the files you name.

Set it up in this order

  1. Pick the job first, then the smallest set of apps it needs. An inbox briefing needs Gmail, a meeting brief needs Google Calendar plus the documents in Drive, an issue investigation needs GitHub.
  2. Install and connect each plugin in ChatGPT under Settings → Plugins, or on your phone in the Dot's profile under Customize → Plugins. Plugin connections are shared across ChatGPT, ChatGPT Work, Codex and your Dot, so an app you already connected carries over. Check which account each one signed in with.
  3. Set a permission level for each app. The options are Always ask, Allow read actions, Allow low-risk actions and Allow all actions (Managing app permissions in ChatGPT). Allow read actions lets it read without asking and asks before any change, which fits a mailbox or Drive you only want summarised. Allow low-risk actions approves low-risk actions on its own, while higher-risk actions may still need your confirmation or be denied. Allow all actions carries elevated risk; the account-wide selector doesn't offer it, and it may appear only for an individual app or connected account.
  4. Add custom rules for anything that leaves your hands. Open Settings → Personalization and select Custom rules under Permissions. Choose Ask before taking action for sending messages to customers and Hand off to you for deleting shared project files, the two examples OpenAI gives (Control your dot). Custom rules don't connect an app and don't override built-in safeguards.
  5. Set up Slack, Teams or texting separately, from the Dot's profile on desktop. A messaging channel is only where you talk to the Dot: adding it to Slack doesn't open your inbox, Drive or computer (Connect computers and apps).
  6. In the first message, name the account, the job and what it may do. Connecting Gmail or adding the Dot to a channel doesn't start a watch, so say what to watch and ask the Dot to confirm the recurring task back to you.
  7. Run the one-minute write test for that app (listed further down), then start the real job.

Where it gets stuck, and what worked

  • "user cancelled MCP tool call" when nobody cancelled. It appeared on Gmail sends and replies (odedd) and in an Outlook cleanup (cedar_circuit), and the message never said whether the action had happened. What worked: look in Sent, the calendar or the repository first, and retry once only if the action is really missing.
  • A yes that doesn't reach the exact write. On GitHub a general go-ahead didn't cover a later batch of files or a new branch, and the batch went through once the question named the files, the private repository and the draft PR (jabdori). A yes typed in the main Dot chat was not accepted by the delegated task (Beowulfwar). What worked: ask the Dot to phrase the approval it needs as one sentence naming the item, the place and the action, and answer it inside the affected task. Asking for a draft never authorises sending it (Getting started with your dot).
  • The approval prompt sits in another conversation. In uehara's three-day test, permission prompts showed up only in a separate conversation, and work stopped silently whenever a prompt went unanswered (uehara). What worked: open the Dot's profile, go to Activity, and answer the waiting request in that task.
  • Reading works, but the watch or the write doesn't. A Gmail watch stopped waking the Dot while an on-demand search still found the email (anotb). What worked: test the write path on something harmless, and ask for one line every evening listing what the watch matched, so a watch that has gone quiet shows up within a day.
  • Two routes into the same app. The GitHub plugin kept working while the Dot's cloud shell had no GitHub login of its own (s520145), and a Google sign-in inside the Dot's browser is separate from the Drive plugin. What worked: ask "which route did you use?", and when one fails, tell the Dot to use the plugin.
  • The wrong account, or a narrower view than you think. Connecting a second Gmail account stopped a new-mail trigger, with no way to choose which account it watched (Daareek), and uehara's summary of today and tomorrow read only the main calendar and missed two events in another one. What worked: have the Dot name the account, calendar or repository in every answer until you trust the setup, and check every watch again after you connect another account.
  • The connection expired. Open the reconnection request, reconnect, then repeat the same small lookup before resuming the wider job.

What it can do, and what it can't yet

It can:

  • read and summarise across the apps you connect, and write exactly as asked: uehara's Dot created and deleted calendar events, sent email on "send it" but not on "draft", and created and deleted Notion pages
  • research your connected apps on its own and suggest ways to help; the tools it uses for that research can't send messages, change app content, or control your browser or computer (Control your dot)
  • keep reading while you're away: it can review connected information proactively and form memories from it, even when you haven't asked about it (Getting started with your dot; case 4 below)

It can't yet, or not reliably:

  • let you choose which Gmail account a new-mail trigger watches (Daareek)
  • connect a work mailbox that runs on hosted IMAP (Gmail guide)
  • get a standalone email address of its own; at launch it uses the personal account you connect
  • forget what it learned when you disconnect an app; only deleting the Dot removes that information
  • undo changes already made in connected apps, or recall messages already delivered, when you delete it (Control your dot)
  • carry one GitHub approval across several different writes (GitHub guide)

One message to start with

I've connected [app] with the account [account]. Before any real work, tell me which account you're using and what you're allowed to do there: read, draft, send or change. Your job is [job]. Never send, post, delete or share anything without first asking me in one sentence that names the item, the place and the action. Watch [source] for [what matters], confirm the watch back to me, and every evening send one line listing what it matched today, even if nothing. Now run one small test: [the write test for this app], read the result back to me, and tell me which route you used.

The five problems behind most app failures

These come from the failure reports on the six app guides (Gmail, Google Calendar, Slack, Google Drive, GitHub and Teams), grouped by what went wrong instead of by app. Each one links to the guide where the reports are told in full. The fixes are in the short version above.

1. "User cancelled" when you cancelled nothing

The message "user cancelled MCP tool call" turned up in Gmail sends and replies, in Google Calendar event creation, in an Outlook cleanup and in GitHub uploads (odedd and cedar_circuit on Gmail, odedd again on Google Calendar, tndd on GitHub). In none of these reports did the message say whether the action had actually happened. Some sends never went out, one calendar create worked on the next try, and one GitHub upload that looked lost turned up hours later (konard). More reports with the same wording, in Slack, a private mail server, a spreadsheet, JobTread and the Dot's own task creation, are collected under Dots not working, section 12.

2. A yes that doesn't reach the exact action

Approvals are matched against the specific write. On GitHub, a general go-ahead didn't cover a later batch of files, a new branch or a comment (jabdori, seeton, carabinshely), a saved merge rule didn't cover the merge (DmitrL-dev), and a yes relayed from the main Dot conversation into a delegated task was treated as untrusted (Beowulfwar, misosiruda). On Gmail, a sign-in code the Dot found in the inbox was refused because it had to come from the user (ammpanda). What unblocked one GitHub case was a single question that named the files, the repository and the draft PR together.

3. Reading works, so the connection looks fine, but writing or watching doesn't

In almost every failure report the read side kept working. A Gmail watch stopped waking the Dot while an on-demand search still found the email (anotb, Daareek on Gmail). A Dot in Slack read channels but couldn't reply, with error -32603 (Slack). Drive search and metadata kept working while a 514 MB upload and a 52 MB download never arrived (COOLak, sbosell on Google Drive). A successful lookup tells you little about a send, an upload or a trigger.

4. Two routes into the same app

Several apps can be reached two ways, and the two fail separately. The Drive plugin uses the permissions you approved in ChatGPT, while a Google sign-in in the Dot's own browser is a separate login that Google challenged (COOLak on Google Drive). The GitHub plugin kept working while gh CLI and git push in the Dot's cloud shell had no login (s520145, konard and f0909172434 on GitHub). In Slack, the Dot's dedicated sending route worked again while the general message route still returned -32603 (nittajun, markguoz on Slack). Teams messaging and the Teams plugin are separate connections too (Teams).

5. The wrong account, or a narrower view than you think

Adding a second Gmail account stopped a new-mail trigger, and the task editor had no way to choose which account it watched (Daareek). A Dot that couldn't read the calendar list worked from the primary calendar only, so its clash check covered one calendar (미루 on Google Calendar). A Dot with its own Workspace account could see whatever that account could see (Nicole Leffer). On GitHub, an upstream repository missing from the app's repository list returned 403 (articultur), and a repository selection that never saved led to refusals (partialobserver). A hosted IMAP mailbox can't be connected at all (TheIgnacioLok on Gmail).

A one-minute write test for each app

Each test checks the write path on something harmless before a real job depends on it.

  • Gmail: "Which Gmail account are you reading? Find the newest email I sent you, tell me its subject and date, then write a draft reply but don't send it. Tell me where the draft is."
  • Google Calendar: "List the calendars you can see. Create a private 15-minute event called 'Dot test' tomorrow at 9 with no guests and no video link, read it back to me, then delete it."
  • Slack: "Send me a DM that says 'Dot test', read it back from Slack, and tell me which sending route you used."
  • Google Drive: "Using the Drive plugin, not the browser, create a one-line Doc called 'Dot test' in my Drive, then open it and read the line back to me."
  • GitHub: "Tell me whether the GitHub app can write to [repository]. Before creating a one-line test file on a new branch, ask me for approval in one sentence that names the file, the branch and the action."
  • Teams: "Is Teams connected as a plugin, as a messaging channel, or both? Summarise my last three unread Teams messages and say which account they came from."

Guides for each app

Open the matching connect guide, then the use case that shows the app at work:

Contact methods are separate from apps

A contact method is where you talk with your Dot; an app connection gives it tools and information. Slack or Teams messaging is set up from the Dot profile, and connecting a messaging channel doesn't grant access to your inbox, other apps or your computer (Meet dots; Connect computers and apps). Plugin permissions are shared across Dots, ChatGPT, ChatGPT Work and Codex (Dots privacy, security, and safety FAQs), and local skills need a connected computer.

Texting your Dot is a limited beta for some Pro users in the United States (Getting started with your dot). It uses a third-party provider, isn't offered in Business or Enterprise workspaces, and may not reach every US Pro account. You can still message or call your Dot in ChatGPT, or connect Slack or Microsoft Teams. Learn's Message your dot page lists texting as coming soon. On Enterprise, phone messaging through iMessage, RCS or WhatsApp is unavailable at launch (Manage dots in ChatGPT workspaces).

Real cases (permissions, connect vs watch, setup, proactive reading)

Each case below shows one boundary worth knowing before you connect an app. Keep sends and public posts behind your yes.

1. Scope what Slack, Gmail and Drive can touch, and how to revoke it (indie)

Bluemediaworld's permissions guide (Alex Kim, Oct 2026) walks the Plugins screen before you connect: a Dot only gets apps you approve; background research is read-only while acting (send, post, edit) goes through your Custom rules; password changes and money transfers always hand off to you. Disconnecting an app stops new access but doesn't erase what the Dot already learned, and resetting the Dot is the full cleanup path. Also check the third-party app's own connected-apps list as a backstop.

What you learn: treat revoking as stopping new access, not as wiping memory.

2. Connecting Slack or Gmail does not create a watch (indie)

Flavio Copes's OpenAI dots deep dive (Sep 30, 2026) separates three connection types: messaging channel, plugin, and your computer. Adding a Dot to #feedback or connecting Slack or Gmail doesn't create a monitoring task by itself; you must say what to watch for and ask the Dot to confirm which events it can follow. Plugins (Gmail, Drive, GitHub) share ChatGPT's Plugins permissions; a messaging channel alone never grants inbox access.

What you learn: "connected" doesn't mean "watching", so assign the source and confirm the watch.

3. Connect apps in setup, manage Plugins, and put sends on ask-before (official)

Meet dots setup offers app connections (and, in the desktop app, computer access); you can start talking while context gathers. Messaging in Slack or Teams is separate from plugins; connecting a channel doesn't connect your inbox or local computer. Getting started with your dot covers reviewing permissions on Customize → Plugins, disconnecting apps, and Custom rules with Ask before taking action or Hand off to you for consequential steps. Introducing dots describes the wider plugins ecosystem.

What you learn: connect during or after setup, then put outbound sends on ask-before before you widen a schedule.

4. A connected Drive gets read without a fresh request (OpenAI Community)

padiwe2753 (17 October 2026) had connected Google Drive and had talked about a thesis with the Dot earlier. Later the Dot read a newer private document without being asked to review it and flagged an unrelated sentence that looked like a copy-paste mistake. The catch was useful, and it still raised the question of what stops the Dot from opening other private documents next. The post asks for per-file or per-folder opt-in, a way to keep on-demand access while turning off proactive reads, and a visible history of which files were read (OpenAI Community). A Dot can review connected information proactively and form memories from it, even when you haven't asked a specific question about it (Getting started with your dot).

What you learn: connecting an app for one task also opens it to proactive reading, so connect an account that holds only what the Dot should see.

More help

Questions about disconnecting an app, approvals that stay blocked or pausing a Dot are answered in a few lines in the ChatGPT Dots FAQ; if your Dot stops with "Your dot is on a break", see what that message means.

Once connections work, jump to scenario playbooks: travel planner, Slack assistant, email triage, meeting prep, morning brief.

Copy this task brief

How to connect apps to a ChatGPT Dot in seven steps, which permission level to pick for each app, a one-minute write test, fixes for the common blockers and one message you can copy.
Download Markdown ↓
Integrations

How to connect Gmail to your ChatGPT Dot: setup, what it reads before you ask, and what to do when a send or mail watch fails (OpenAI Dots)

Connect the Gmail plugin and your ChatGPT Dot can search and read that account's mail and flag what you missed, often before you ask. Setup steps, what it reads on its own, how to keep sending in your hands, fixes for cancelled sends and quiet mail watches, and a copyable first message.

Integrations

How to connect Google Calendar to your ChatGPT Dot: setup, clash checks, and what to do when an edit fails or it checks only one calendar (OpenAI Dots)

Connect the Google Calendar plugin and your ChatGPT Dot can read your schedule, flag clashes it finds in your mail and chat, and create or move events when you allow writes. Setup steps, which custom rule stops a delete, fixes for Bad Request errors, stray Meet links and one-calendar checks, and a copyable first message.

Integrations

How to connect Google Drive to your ChatGPT Dot: setup, what it reads on its own, and what to do when a Sheet breaks or an upload is refused (OpenAI Dots)

Connect the Google Drive plugin in ChatGPT and your Dot can find, read and update Docs and Sheets with the same permissions. Setup, what it reads without being asked, fixes for broken Sheets, stalled transfers and refused uploads, and prompts to copy.

Integrations

How to connect Notion to your ChatGPT Dot: setup, board checks and 14 real Notion cases (OpenAI Dots)

Connect the Notion plugin in ChatGPT and your Dot reads, live, the pages that account can open, and creates or edits pages when the connection allows. 14 real cases, from a reporter's schedule checked for clashes within a minute to a Sprint Board worked every 15 minutes, and a delete that needed Chrome.

Integrations

How to add your ChatGPT Dot to Slack: setup, admin approval, reply times, and what to do when replies never arrive (OpenAI Dots)

Open your Dot's profile, choose Add and connect Slack, then DM it or mention it in a thread; on Enterprise an admin must allow dots in Slack first. Mention it again in every thread reply, test in a DM first, and don't remove the connection to rename it. Real ChatGPT Dots (OpenAI Dots) Slack setups, MCP -32603 reply failures and the fixes that worked.

Integrations

How to connect a ChatGPT Dot to Microsoft Teams: setup, admin approval, chat vs plugin and 7 real cases (OpenAI Dots)

Open your Dot's profile, select Add and choose Microsoft Teams; on Enterprise an admin has to allow it first. Reading your Teams chats, Planner and meeting transcripts needs the separate Teams plugin. 7 real cases with Teams, Outlook and Microsoft 365.

Integrations

How to connect GitHub to your ChatGPT Dot: setup, PR reviews, stuck writes and 18 real repo cases (OpenAI Dots)

Install the ChatGPT app on GitHub for only the repos your Dot should see; it reads code on demand and hands fixes to Codex. 18 real cases, from a CI break caught before macos-14 runners retired to a read-only map of 415 files across three mixed systems, plus public bug reports on approvals, big files and the cloud shell.

Sources & further reading