Browser Automation Best Practices

Reference documentation for browser automation using the Claude in Chrome MCP tools.

---

Core Principles

1. User drives, agent assists - User often completes actions faster than agent navigation
2. Capture confirmations - Always screenshot after user completes bookings/forms
3. Extract data proactively - Pull addresses, phone numbers, confirmation numbers from pages
4. Enrich tasks - Add captured details to Things tasks for future reference

---

Confirmation Capture

When the user says they completed something in the browser (booking, form submission, purchase):

Immediately screenshot to capture details:

mcp__claude-in-chrome__computer(action="screenshot", tabId=<id>)

Then extract key information:

• Confirmation/reservation numbers
• Dates and times
• Addresses and phone numbers
• Reference links or order IDs

Add to relevant task notes:

mcp__things__update_todo(id, notes="[existing]

---
Booked: [date/time]
Confirmation #: [number]
Location: [address]
Phone: [phone]")

Trigger Phrases

Screenshot when user says:

• "I booked it"
• "Done"
• "I submitted it"
• "Order placed"
• "Signed up"
• "Scheduled"

---

Tab Management

Session Start

Always get tab context first:

mcp__claude-in-chrome__tabs_context_mcp(createIfEmpty=true)

Creating Tabs

Create new tabs for new tasks instead of reusing:

mcp__claude-in-chrome__tabs_create_mcp()

Tab ID Errors

If a tool returns "tab doesn't exist":

1. Call tabs_context_mcp to get fresh tab IDs
2. Use a valid tab ID from the response

---

Navigation Patterns

Waiting for Page Load

After navigation, wait before interacting:

mcp__claude-in-chrome__computer(action="wait", duration=2, tabId=<id>)
mcp__claude-in-chrome__computer(action="screenshot", tabId=<id>)

Cookie Banners

Dismiss cookie banners for privacy (click "Deny" or "Reject"):

mcp__claude-in-chrome__find(query="reject cookies button", tabId=<id>)
mcp__claude-in-chrome__computer(action="left_click", ref="<ref_id>", tabId=<id>)

Form Filling

Use form_input for form fields:

mcp__claude-in-chrome__form_input(ref="<ref_id>", value="text", tabId=<id>)

---

Handoff Patterns

User Takes Over

When user can complete faster (login, complex forms):

1. Navigate to the right page
2. Screenshot and explain what's needed
3. Say: "I've navigated to [page]. You can [complete action] - let me know when done and I'll capture the confirmation."
4. When user confirms, screenshot and extract details

Portal/Login Walls

If a site requires login:

1. Navigate to the login page
2. Screenshot
3. Ask: "This requires login. Want me to continue after you sign in, or take it from here?"

---

Data Extraction

Reading Page Content

For text-heavy pages:

mcp__claude-in-chrome__get_page_text(tabId=<id>)

Finding Elements

For specific elements:

mcp__claude-in-chrome__find(query="confirmation number", tabId=<id>)
mcp__claude-in-chrome__read_page(tabId=<id>, ref_id="<ref>")

---

Error Handling

When Things Go Wrong

If browser tools fail 2-3 times:

1. Stop attempting
2. Report what happened
3. Offer alternatives (user takes over, try different approach)

Common Issues

Issue	Solution
Page not loading	Wait longer, refresh, check URL
Element not found	Try different query, screenshot to verify
Click not working	Verify coordinates with screenshot, try ref instead
Tab ID invalid	Call tabs_context_mcp for fresh IDs

---

Anti-Patterns

DON'T guess at page state

Always screenshot before making assumptions about what's on screen.

DON'T rapid-fire actions

Wait between actions for page to respond.

DON'T fight with complex forms

If a form is complex or requires login, hand off to user.

DON'T forget to capture

After user completes actions, always screenshot to extract confirmation details.