To revoke an AI agent's access properly, do three things: remove it on the account side (Google, Microsoft or whichever service it connects to), disconnect it inside the agent, and clear what it remembers. The account side is the reliable off switch, because it cancels the token the agent uses. But removing access only stops future reading and sending; it does not delete what the agent already copied or learned. This guide covers each step for the major agents, based on official docs as of November 2026.

Why revoke from both sides?

When you connect an agent to Gmail or Outlook, the account issues it a token, a digital key with certain permissions. The agent stores that key.

  • Revoking on the account side cancels the key. Even if the agent or its vendor still has a copy, it no longer works.
  • Disconnecting in the agent removes the connector from its settings, so tasks stop trying to use it and you are not prompted to reconnect.
  • Clearing memory removes facts the agent extracted while it had access.

How do I remove an agent from my Google account?

Google's help page Manage links between your Google Account and apps from other developers gives these steps:

  1. Open the pageGo to myaccount.google.com/linkedapps, signed in to the right account.
  2. Choose the listSelect Access to your Google Account.
  3. Find the agentSelect the app, such as the agent's connector or the vendor's name.
  4. ReviewSelect See details to check what access it had.
  5. RemoveSelect Remove access, then Confirm.

Google says that afterwards "the app can't access your Google Account". It does not say the app deletes data it already has. Separately, if you used "Sign in with Google" for the agent, removing that sign-in link "doesn't delete your data on the app".

If you see several entries for the same vendor (for example, one for the agent and one for a related product), check each. OpenAI says plugin permissions are shared across ChatGPT, ChatGPT Work and Codex, so a connection made in one may appear once but serve all three.

How do I remove an agent from my Microsoft account?

For a personal Microsoft account (Outlook.com, Hotmail, OneDrive), Microsoft lets you review apps that have access to your data and remove their permissions. Third-party guides, such as OneCal's walkthrough, point to account.live.com/consent/Manage. We could not retrieve Microsoft's own help page to confirm the current path, so if that address does not work, sign in at account.microsoft.com and look under privacy or security for apps and services with access.

  1. Sign in to your Microsoft account.
  2. Open the list of apps with access to your data.
  3. Select the agent's app and choose to remove its permissions.

Work or school Microsoft 365 accounts are different: an administrator controls app consent. Ask your IT team. Claude Duet did not support Microsoft 365 at launch.

How do I disconnect inside each agent?

Agent Where to disconnect Does disconnecting erase what it learned?
Meta Muse Connector settings; Meta says you can disconnect at any time Not documented; tell Muse to "forget", no retention period published
OpenAI dots Connected apps (shared with ChatGPT and Codex) No; reset or delete the dot
Claude Duet Switch the app off in Claude connectors; this also switches it off in Duet No; delete or reset Duet
Grok Bot Bot settings and the shared cloud computer's logins Sign out of sites on the cloud computer; deleting a single Bot is a soft delete
OpenClaw Your config: remove the integration's credentials, rotate tokens Delete local memory and session files yourself

Based on vendor documentation and launch coverage, as of November 2026.

Meta Muse

Meta says you can disconnect any connector at any time (launch post). We could not find Meta documentation with exact steps for removing an individual perpetual approval, so if you cannot find one, disconnect the connector. If you enabled the Mac Messages connector, also remove Full Disk Access for Muse in macOS System Settings → Privacy & Security.

OpenAI dots

Disconnect apps from the dot's settings and review connected apps in ChatGPT and Codex too. OpenAI says disconnecting does not remove what the dot learned, and memory cannot be edited item by item, so reset or delete the dot for a clean slate. Data requests go through OpenAI's Privacy Portal or dsar@openai.com.

Claude Duet

Anthropic says connectors carry over from Claude, and switching one off in Claude switches it off in Duet (safety article). Duet keeps what it learned until you delete or reset it.

Grok Bot

All your Bots share one cloud computer with its browser sessions and logins, per xAI's FAQ. Revoke the account-side token and also sign out of the service on the cloud computer. Deleting a single Bot is a soft delete; deleting Grok Bot entirely deletes your Cursor account, with data removed within 30 days.

OpenClaw

Remove the integration's credentials from your configuration, rotate any gateway tokens, and remove DM pairings you no longer need. Then run openclaw security audit. See OpenClaw security hardening.

How do I clear what the agent remembers?

Revoking access does not reach backwards. To remove what the agent already knows:

  • Reset or delete the agent where the vendor offers it (dots, Duet).
  • Ask the agent to forget specific things where supported (Muse).
  • Submit a data deletion request to the vendor.

Our guide to deleting AI agent memory covers each, and how long AI agents keep data explains retention.

What if I think access was misused?

  1. Revoke access on the account side first.
  2. Change that account's password from the service's own site and review sign-in activity, sent mail and forwarding rules.
  3. Save the agent's activity log.
  4. If money moved, contact your card issuer or bank promptly. In the US, report fraud at reportfraud.ftc.gov or IC3.
  5. Report the incident to the agent vendor's support.

See what to do if your AI agent did something you didn't ask for a fuller checklist.

Bottom line

Revoke on the account side, disconnect in the agent, then clear memory. Make it a habit when a task ends, not only when something goes wrong; that is least privilege in practice. If you are leaving an agent entirely, our sister site explains how to cancel a personal AI agent.