ChatGPT Chrome Setup: Six Steps to Minimal Permissions

<!-- chatgpt-chrome-workmode-setup-sop | sop | ChatGPT Chrome Setup: Six Steps to Minimal Permissions -->

On October 17, OpenAI shipped a sweeping update to the ChatGPT extension for Chrome: it can now read and operate your open tabs, fill forms across pages, compare prices across sites, and run multi-step tasks, with browser conversation threads syncing back into the desktop app. If your mental model still says ChatGPT cannot operate a browser, it is out of date: per the official documentation, a paid account, the desktop app, and one Chromium-based browser are all it takes. But a full chapter of permission settings sits between "it can act" and "you can trust it to act" - and the official docs themselves flag two high-risk options. This SOP breaks the setup into six steps: prerequisites, installing the extension, invoking it, choosing among three entry points, tightening permissions to the minimum, and the security habits that make the configuration durable. Every operational detail comes from learn.chatgpt.com; information is current as of October 2026, and the live pages are authoritative.

Step 1: Get the Prerequisites Right - Three Things, No Substitutes

Check three items before touching any settings:

1. A paid ChatGPT plan. Officially every paid tier works; the free tier does not get this capability. Plus, Pro, and team users can move on. Free users have no path here and no official workaround.
2. The ChatGPT desktop app. The installation entry point lives in the desktop app, not the web UI, on both Windows and macOS. Install it first - it is the first stop on the chain.
3. One Chromium-based browser. Five are supported: Chrome, Edge, Brave, Opera, and Vivaldi. Only these five. Firefox and Safari are not supported - not a version problem but a hard boundary in the product, so stop looking for a toggle that does not exist.

macOS users have one extra layer: Computer Use requires two system permissions, Screen Recording and Accessibility. The system will ask for them after the extension is installed; approve each one.

Expected result: a paid account signed into the desktop app and one of the five browsers chosen as primary. Only then move to Step 2. One selection note: the official docs mark no capability differences among the five, so pick by habit - if Chrome is already your daily driver, do not migrate to Edge for this feature.

Step 2: Install the Extension - a Six-Node Chain

The official install path is a chain. Walk it in order without skipping:

1. Open the ChatGPT desktop app and go to Settings > Computer Use;
2. Select your browser from the list and click Install;
3. The system jumps to that browser's extension store;
4. Approve the permissions the extension requests on the store page;
5. Return to the Computer Use settings page in the desktop app;
6. Confirm the browser now shows Manage.

The last node is the one people skip most: they approve the store dialog, close the window, and later assume the extension failed. The pass criterion is simple - the word Manage next to your browser in Computer Use settings. No Manage means the chain is broken somewhere; restart from node 2. On macOS, the system will additionally request Screen Recording and Accessibility mid-chain; that is normal flow, not a glitch to dismiss.

One clarification: the permissions approved on the store page and the site-level permissions in Step 5 are different things - the latter are asked per hostname each time ChatGPT wants to act on a domain. Approving a long list during installation is not a pass for every website.

Expected result: Manage appears next to the browser you selected, meaning the extension is under the app's control.

Step 3: Invoke ChatGPT - Three Ways In

Once installed, there are three ways to bring ChatGPT into a task. Pick by context:

Way one: @ mention in a conversation. In a ChatGPT Work or Codex conversation, type @Chrome, @Edge, @Brave Browser, @Opera, or @Vivaldi to pull that browser into the task. The official example reads "@Chrome open Salesforce and update the account from these call notes" - one line naming the target app, the action, and the source material. If the browser is not running, ChatGPT can open Chrome itself.

Way two: the Chrome sidebar. On macOS, press Cmd+Shift+. (period) to summon the sidebar chat, or right-click any page and choose Ask ChatGPT. From the sidebar you can attach the current tab or selected text as context. Its value is not switching apps: while reading a long article or filling a complex form, you ask about the page in place, with the tab and highlights carried in as context automatically. Officially, all of this content - page content, selected text - is labeled untrusted context; more in Step 6.

Way three: Appshot. Press Command twice in a row (Cmd+Cmd) to attach the frontmost Mac window to the conversation as a screenshot plus readable text - built for "how do I deal with what is on my screen right now."

One behavioral detail: browser tasks run inside Chrome tab groups, so pages belonging to the same task stay collected and a multi-step task does not scatter tabs across your window.

Expected result: all three routes bring ChatGPT onto the page you are looking at, with task pages collected into tab groups.

Step 4: Choose Among Three Entry Points - the Extension Is Not the Only Door

Officially there are three entry points: @Computer (the desktop app itself), @Browser (the built-in browser inside the desktop app), and @Chrome (your daily, logged-in browser). They map to three task types, and the official selection logic can be copied verbatim:

• Tasks that need your existing login state - SSO intranets, CRM, work email, and logged-in sites like LinkedIn, Salesforce, Gmail, or internal tools - go to the extension (@Chrome). It runs in your daily browser with your sessions already in place; that is its irreplaceable use case.
• Tasks against localhost, preview pages, or anything where a separate configuration does not matter - use @Browser, the desktop app's isolated built-in browser.
• Services with a dedicated plugin - use the plugin first instead of forcing a browser agent.

One-line decision rule: login state in the browser means extension, local pages mean @Browser, plugin exists means plugin. Treating the extension as one of three doors rather than the only door is both an efficiency choice and a permission choice - the extension can see the browser's full context, including text you have highlighted, so there is no reason to let it into tasks that do not need it. The officially listed strengths help decide what to hand over: researching markets and comparing sources, refining Google Workspace or Microsoft 365 files, multi-step operations across pages, filling forms, and logged-in tasks on LinkedIn, Salesforce, Gmail, or internal tools. Tasks outside that list deserve a second thought.

Expected result: each task type has a home, and the extension only handles the slice that genuinely requires your login state.

Step 5: Tighten Permissions - How to Click the Four Levels

This is the core of the SOP. New sites are ask-first by default: when ChatGPT wants to act on a site, it prompts first, keyed to the hostname (example.com). Each prompt offers four levels:

• Allow once: permit this action only; it expires immediately after, and the next action prompts again;
• Allow for this site: standing permission for that one site;
• Allow for all sites: permission for every site;
• Decline: refuse.

The daily-driving strategy is Allow once: one extra click buys you eyes on every step. Reserve Allow for this site for a small set of high-frequency trusted sites, such as the CRM it operates daily - if you would happily re-confirm that site twenty times a week, stay on Allow once; upgrade to site-level permission only when the site is trusted and the prompting gets tedious. Allow for all sites is explicitly flagged as high-risk in the official documentation - it expands the permission surface to every domain, and any single induced page would hold an action permit. This SOP's position: do not use it. Decline costs nothing; when in doubt, refuse, and use Allow once after you have thought it through.

Permission management lives in Settings > Computer Use under Manage, next to each browser: a domain allowlist (stop asking) and a blocklist (forbidden), both editable anytime. Note one easily missed fact: these site permissions are shared across all five browsers - a domain you allow in Chrome is equally allowed in Edge, so think in terms of the whole fleet when granting or revoking.

The second officially flagged high-risk item is browser history access. With it enabled, ChatGPT can read your browsing history - another significant expansion of the permission surface. Unless you have a concrete use case, keep it off.

Expected result: an allowlist of a few domains you truly trust, Allow for all sites never clicked, history access off.

Step 6: Security Habits - the Other Half of the Story

Permissions govern what it can do; habits govern what information it acts on. The official documentation states the principle bluntly: treat page content, selected text, and video transcripts as untrusted context. Unpacked, that is three habits:

1. Page content lies. A crafted page can steer an agent into wrong actions, which is why the official guidance is to review the page before letting ChatGPT act. The few seconds when a permission prompt pops up are not for hurrying to Allow - glance at what it intends to do and on which domain.
2. No standing permissions for the agent. Echoing Step 5: Allow once is the default gesture, the allowlist stays short, and Allow for all sites stays untouched.
3. Local file access has a premise. The extension can reach the local file system and installed plugins, but only after you explicitly authorize Computer Use from the desktop app - it does not quietly read your disk. The flip side: once granted, that authorization is a wide surface, which is why the store-permission review in Step 2 deserves real attention. One often-missed detail: threads started in the browser sync into the desktop app, so browser task records do not stay in the browser; factor that in before running tasks on sensitive pages.

One note on video tasks: YouTube videos can be explained or summarized through timestamped subtitle transcripts, and those transcripts are also untrusted context - confirm the source video is trustworthy before acting on one.

Expected result: you review the page before every action, standing permissions are zero or near zero, and local file access is a deliberate decision rather than a habit of clicking Allow.

Appendix: Known Gotchas

• Cannot find the install entry. It is in the desktop app under Settings > Computer Use - not in web settings, and not found by searching the extension store. Install the desktop app first.
• Firefox/Safari users stuck at Step 1. Only five Chromium browsers are supported. Switch browsers; there is no setting to change.
• Opera has no sidebar chat. By design: Opera lacks the sidebar chat but supports desktop workflows. For the sidebar, use Chrome, Edge, Brave, or Vivaldi.
• Installed but nothing changed. Nine times out of ten the chain's last node was skipped - confirm Manage in Computer Use settings; otherwise redo Install.
• Permissions leaking across browsers. Site permissions are shared by all five browsers; granting to Chrome grants to the fleet, and so does revoking.
• Still on Atlas. Per The Decoder and AIBase reporting, OpenAI has said it will gradually wind down the standalone Atlas browser and share migration guidance toward the extension - this SOP is that direction; there is no Atlas update to wait for.

FAQ

Q1: Which browsers are supported?

A1: Five, all Chromium-based: Chrome, Edge, Brave, Opera, and Vivaldi. Firefox and Safari are not supported. Opera lacks the sidebar chat but works for desktop workflows; the other four support the sidebar, summoned on macOS with Cmd+Shift+.

Q2: Can free-tier users do this?

A2: No. Officially, all paid ChatGPT tiers get the capability; the free tier has neither Computer Use nor the browser extension, and there is no workaround.

Q3: Will my browsing data be stored?

A3: Two things are confirmable from the official docs: conversation threads started in the browser sync into the desktop app and can reference each other, and site permissions are recorded per hostname in the allowlist/blocklist, shared across all five browsers. For retention of conversation and browsing data, defer to OpenAI's live privacy terms - this article makes no promises on the company's behalf.

Q4: How do I revoke permission for a website?

A4: Go to Settings > Computer Use, click Manage next to the browser, and delete the domain from the allowlist; the blocklist can ban a domain outright. Allow once grants never become standing permissions. The list is shared across all five browsers - revocation is fleet-wide too.

Q5: What does this have to do with the Atlas browser?

A5: Atlas was OpenAI's own AI browser, previously developed as a standalone product. Per The Decoder and AIBase reporting, OpenAI has said it will gradually discontinue the standalone Atlas and share migration guidance toward the ChatGPT extension - the exact setup in this article. The strategic logic: building on Atlas's experience, embed into Chrome's roughly sixty percent market share instead of competing from zero.

---

Sources

• learn.chatgpt.com official documentation: entry points, supported browsers, installation chain, permission levels plus allowlist/blocklist, macOS permissions, untrusted-context principle, Appshot and sidebar, tab-group behavior
• The Decoder (October 17, 2026) and AIBase (citing OpenAI): the October 17 extension update, paid-tier availability, phasing out Atlas with migration guidance
• aicatchup: extension evolution timeline, May to October 2026

Related reading

• Self-host the open-source Dots in three commands: Self-host the open-source Dots in three commands
• AI Agent Tool Calling SOP: Teaching Large Models to Call Tools: AI Agent Tool Calling SOP
• Claude Code Subagents: A Five-Step SOP for Cheaper AI Coding: Claude Code Subagents: a five-step SOP