Your AI can now click, scroll, and type. It fills forms, navigates menus, screenshots bugs, and even shops for you. Here’s what that actually looks like in practice — four workflows you can copy today, no code required.
What Is Codex Browser Automation? (And Why It’s Different)
Codex is ChatGPT’s desktop app (Mac and Windows). It has two superpowers: browser automation and computer use. It can control a browser like a human — click, type, scroll, navigate, screenshot, run JavaScript. But the magic is in three command prefixes that change everything:
@browser vs @Chrome vs @computer — Which Prefix When
| Prefix | Scope | Use When |
|---|---|---|
@browser |
Embedded browser (clean, isolated session) | Testing public sites, QA, no login needed |
@Chrome |
Your active Chrome window (your logins, cookies, state) | LinkedIn, Gmail, any authenticated site |
@computer |
Full desktop (apps, OS, file system, iPhone mirroring) | Desktop apps, system tasks, iPhone control |
Critical: @Chrome sees your real logged-in session. @browser starts fresh every time. @computer is powerful but risky — use a VM or secondary machine.
What You Need (2-Minute Setup)
- ChatGPT Plus/Pro/Team/Enterprise — Codex requires a paid plan
- Codex desktop app — Download from chatgpt.com/desktop or Mac App Store
- Enable browser access — Codex Settings → General → “Browser access” → On
- For @Chrome: Have Chrome open and logged into target sites
That’s it. You’re ready.
Workflow 1: Automated QA Testing That Finds Bugs Humans Miss
The Problem
Human QA follows the happy path. Fill form correctly → click continue → confirm success → move on. We don’t intentionally break things. We don’t test every viewport. We don’t have time.
The Prompt (Copy This)
@browser to test the open onboarding flow for usability and mobile responsiveness.
Please screenshot as you go, make a Google spreadsheet of any issues you identify.
What the Agent Actually Does
- Resizes viewports: mobile, tablet, desktop, ultrawide
- Opens every menu, toggles every dropdown, clicks every button
- Repeatedly moves through the entire flow across layouts
- Screenshots every step
- Documents issues in a Google Sheet with reproduction steps
The Result: 11 Issues in a Google Sheet with Repro Steps
In the source case (ChatPRD’s onboarding flow), the agent found:
– High-severity: “Continue” button clickable without required selection → user trapped in stalled state
– Layout: Overflow on mobile, touch targets too small
– Accessibility: Missing labels, contrast issues
– Responsive: Breakpoints behaving unexpectedly
Structured evidence trail after an exhaustive pass you realistically wouldn’t complete manually. Review findings → prioritize → decide what’s worth fixing.
Why It Beats Human QA
“The browser agent spent far more time exploring edge cases and failure states. Human QA often follows the intended route: fill out the form correctly, click continue, confirm the success state, move on.”
The agent doesn’t get bored. It doesn’t assume the happy path. It tries to break things because that’s the assignment.
Workflow 2: Persona-Based User Research (Simulate Real Customers)
The Problem
Scripted QA checks whether buttons work. Persona testing reveals where momentum breaks for real users attempting real goals.
The Prompt (Three Personas, One Command)
Dictate this (voice input works great for longer prompts):
I want you to go through the core of my app's chat and document creation mode as three different personas:
1. A product manager trying to quickly create a PRD for a feature they just got out of a meeting about
2. An engineer taking that PRD and trying to make a technical spec out of it
3. A team leader trying to figure out how their team is using the product
Create a doc providing research-style critique about the overall experience, including friction points, places of delight, and improvements for those three personas.
What the Engineer Persona Found (Structural Break)
The engineer persona attempted to reference the document created by the PM persona and hit a “structural break” — cross-thread document references didn’t work. The team knew this limitation existed. But seeing it emerge naturally inside a realistic workflow made the weakness feel consequential.
What the Team Lead Persona Found (Stalled Loading State)
A confusing loading state after an error. The system felt stalled without enough feedback about what was happening. Scripted QA would miss this — the button “works,” the error “displays.” But the experience fails.
Why Roles Beat Checklists
“Role-based testing over scripted QA. The agent is not just checking whether buttons work. It is trying to accomplish a goal inside a believable context, then explaining where momentum breaks down or confidence disappears for the user.”
Goal-driven exploration finds friction that checkbox-driven verification misses.
Workflow 3: LinkedIn Inbox Triage (Delegate the Boring Part)
The Problem
LinkedIn inboxes: high-value messages buried in low-value sorting work. The messages matter. The repetitive navigation doesn’t.
The Prompt (Lightweight Reply Rules)
Please use browser to go through my recent unread messages and reply to any that are super critical for my business or podcast. Leave me notes here on the others about how I should reply. If this is just a friendly thank you note reply in a friendly way.
How It Works
- Open LinkedIn in Chrome (logged in)
- Run prompt with
@Chrome(uses your active session) - Agent works through unread messages
- Replies to simple thank-yous
- Flags conversations needing context/nuance
- Leaves notes for your review
Model Selection Tip: Don’t Overpay for Triage
Started with GPT-5.6 Soul on high effort → overkill. Switched to a medium-effort model → faster, cheaper, same triage quality.
Browser automation adds execution overhead. Match the model to the actual judgment required.
Safety: Human Reviews Everything Important
The agent handles the mechanical sorting. You handle the judgment calls. Nothing sensitive goes out without your eyes on it.
Full implementation guide: chatprd.ai/how-i-ai/workflows/automate-linkedin-inbox-triage-with-ai-browser-automation
Workflow 4: Smart Shopping with Constraints (Fill Your Cart)
The Problem
Sale browsing is time-consuming, low-judgment navigation. Pages of items, filters that reset, sizes that sell out.
The Prompt (Outcome + Context, Not Micromanagement)
@Chrome I'm going to Hawaii this weekend and I need new clothes. Free People has 25% off their sale. Pick out 10 items in size medium that work for Hawaii weather in late July. I'm a breastfeeding mom — comfortable, breastfeeding-friendly, weather-appropriate. End up with 10 things in my cart I can pick from and buy.
What Happened
- Agent searched sale inventory, filtered by size (medium)
- Evaluated styles against constraints (Hawaii weather, breastfeeding-friendly, comfortable)
- Added 10 items to cart (loose pants, maxi dresses, button-down tops)
- Paused at CAPTCHA → human completes verification → hands back control
The CAPTCHA Handoff (Correct Division of Labor)
“That handoff actually felt like the correct division of labor. The agent handled the long and boring part: browsing dozens of listings, checking fit constraints, narrowing options down. I handled the one step specifically designed to require a real person.”
Why Constraint-Based Prompting Works
“Described the outcome and context clearly without micromanaging every search step.”
Give the agent the what and why, not the how. Let it figure out the navigation.
3 Smaller Browser Tasks That Add Up Fast
Repetitive Forms
Camp registrations, procurement forms, insurance paperwork, school permissions. Agent fills, you review, submit.
iPhone Mirroring (Router Config from Hotel)
Claire was traveling, needed to open firewall ports on her home router for SSH access to Mac Minis. The router app was on her iPhone in California. Codex opened iPhone Mirroring, updated router settings, completed SSH setup, closed ports. Nearly impossible manually from a hotel room.
Documents/Spreadsheets (When API Fails)
If a direct integration fails, ask the agent to open the web app and complete the work in the browser. Google Sheets, Notion, Airtable — anything with a web UI.
Safety First: Boundaries Every Beginner Must Set
- Never use
@computeron your primary machine — Use a VM, Windows Sandbox, or secondary device. Full desktop control is powerful and risky. - Watch first runs — Monitor the agent’s actions in the Codex side panel before trusting it unsupervised.
- No financial actions — Don’t let the agent click “Buy,” “Confirm Payment,” “Subscribe,” or “Transfer.”
- No destructive actions — No “Delete Account,” “Factory Reset,” “Remove All Data,” “Purge.”
- Use
@browserfor public sites — Isolated session, no access to your real accounts or cookies. - Break long tasks into check-ins — Agent times out on marathon sessions. Prompt in chunks.
- Keep credentials out of prompts — Never paste passwords, API keys, or tokens. The agent uses your active session via
@Chrome.
Model Selection: Match Effort to Task (Save Money)
| Task Type | Judgment Required | Recommended Effort | Why |
|---|---|---|---|
| QA Testing | High (spot subtle bugs) | High | Needs reasoning for edge cases |
| Persona Research | High (simulate realistic behavior) | High | Nuanced roleplay |
| LinkedIn Triage | Low-Medium (sort, simple replies) | Medium | Pattern matching, not deep reasoning |
| Form Filling | Low (follow instructions) | Low-Medium | Repetitive, structured |
| Shopping | Medium (apply constraints) | Medium | Judgment on fit/style |
| iPhone Mirroring | Low (navigate known app) | Low | Mechanical steps |
Rule: Browser automation already adds overhead. Don’t compound it with an overpowered model for simple tasks.
Codex vs Zapier/Make: Why Browser Automation Wins for Dynamic Sites
| Feature | Traditional Automation (Zapier/Make) | Codex Browser Automation |
|---|---|---|
| Setup time | Hours (API keys, webhooks, field mapping) | Minutes (natural language prompt) |
| Authenticated sites | Complex (OAuth, cookies, token refresh) | Native (@Chrome uses your session) |
| Visual judgment | ❌ (no eyes) | ✅ (screenshots, reads UI) |
| Dynamic sites (React/SPA) | Brittle selectors break on UI changes | Works (interacts like human) |
| CAPTCHA handling | Manual intervention | Manual intervention (same) |
| Cost | $20-100+/mo | Included in ChatGPT Plus ($20) |
| Flexibility | Rigid workflows | Adaptive (agent “figures it out”) |
| Learning curve | Medium (visual builder) | Low (chat interface) |
Traditional automation excels at structured, API-driven workflows. Browser automation excels at the messy, visual, authenticated web that APIs don’t reach.
Your First Workflow: Pick One, Try It This Week
Don’t try all four. Pick the one that makes you think “I waste time on that.”
Hate manual QA? → Workflow 1. Point @browser at your staging site. Get a Google Sheet of bugs.
Drowning in LinkedIn? → Workflow 3. Open LinkedIn in Chrome. Run the triage prompt. Review notes in 5 minutes.
Curious about persona testing? → Workflow 2. Pick three user types for your product. See what the agent discovers.
Sale season coming up? → Workflow 4. Give @Chrome your constraints. Come back to a curated cart.
The setup is two minutes. The payoff is hours back every week. And you’re not just saving time — you’re learning how to delegate to agents, which is the skill of the next five years.
Codex desktop app: chatgpt.com/desktop
OpenAI announcement: openai.com/index/introducing-codex/
Source article: chatprd.ai/how-i-ai/4-hands-free-workflows-using-codex-browser-automation
LinkedIn workflow: chatprd.ai/how-i-ai/workflows/automate-linkedin-inbox-triage-with-ai-browser-automation
Persona workflow: chatprd.ai/how-i-ai/workflows/conduct-ai-powered-user-research-by-impersonating-personas
QA workflow: chatprd.ai/how-i-ai/workflows/automate-web-app-qa-testing-with-ai-browser-automation
