Browser Use: Cloud Browser Automation With Persistent Login Profiles

With over 21,000 downloads and 50 stars, the browser-use skill by @ShawnPana is ClawHub's leading browser automation integration. It connects Claude to the Browser Use Cloud API — a service that provisions cloud browsers on demand, maintains persistent login profiles, and runs autonomous browser tasks using a purpose-built LLM optimized for web interaction.

The Problem It Solves

Browser automation has always required you to own the infrastructure: install Playwright or Selenium, manage browser binaries, handle headless rendering, write CSS selectors or XPath, and deal with sites that detect and block automation. That's before you add login persistence — keeping sessions alive across runs typically means serializing cookies to disk and hoping the site doesn't invalidate them.

The browser-use skill eliminates the infrastructure layer entirely. Browsers run in the cloud. Profiles store cookies server-side and persist across sessions. And instead of writing selectors, you write natural language: "Go to amazon.com and find the price of the MacBook Air M3." A purpose-built browser LLM handles the interaction details.

Architecture

The skill operates in two distinct modes that serve different needs:

Mode 1 — Browser Sessions: Spin up a cloud browser, get a CDP (Chrome DevTools Protocol) URL, and connect Claude directly to it. Claude controls the browser interactively, step by step, with full visibility into what's happening.

Mode 2 — Task Subagents: Fire-and-forget automation. Give a prompt, get back a result. A separate browser-use-llm agent handles all the browser interactions autonomously, returning structured output when done.

Both modes support profiles — persistent containers that hold cookies and login state — so authenticated sessions (Gmail, LinkedIn, Shopify, any logged-in site) survive across runs without re-authentication.

Mode 1: Interactive Browser Sessions

Spin Up a Session

# With profile (recommended — keeps you logged in)
curl -X POST "https://api.browser-use.com/api/v2/browsers" \
  -H "X-Browser-Use-API-Key: $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"profileId": "<profile-uuid>", "timeout": 60}'

Response includes a cdpUrl — the Chrome DevTools Protocol endpoint that Claude uses to control the browser:

{
  "id": "session-uuid",
  "cdpUrl": "https://<id>.cdp2.browser-use.com",
  "liveUrl": "https://...",
  "status": "active"
}

Connect Claude to the Browser

gateway config.patch '{"browser":{"profiles":{"browseruse":{"cdpUrl":"<cdpUrl>"}}}}'

From this point, Claude can use its browser tool with profile=browseruse to navigate pages, click elements, fill forms, and extract content — exactly as if it had a local browser.

Stop a Session (Refund Unused Time)

curl -X PATCH "https://api.browser-use.com/api/v2/browsers/<session-id>" \
  -H "X-Browser-Use-API-Key: $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"status": "stopped"}'

Sessions are billed per minute, and unused time is refunded when you stop early. Max session length is 4 hours.

Mode 2: Autonomous Task Subagents

For tasks that don't need Claude's supervision — extracting data, filling forms, completing purchases — use task subagents:

curl -X POST "https://api.browser-use.com/api/v2/tasks" \
  -H "X-Browser-Use-API-Key: $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "task": "Go to amazon.com and find the price of the MacBook Air M3",
    "llm": "browser-use-llm"
  }'

Always use browser-use-llm — it's a purpose-built model optimized for browser interactions, 3–5x faster than general-purpose models for this task.

Poll for results:

curl "https://api.browser-use.com/api/v2/tasks/<task-id>" \
  -H "X-Browser-Use-API-Key: $API_KEY"

{
  "status": "finished",
  "output": "The MacBook Air M3 is priced at $1,099",
  "isSuccess": true,
  "cost": "0.02"
}

Task Options

Profiles: Persistent Login State

Profiles solve the hardest problem in browser automation: authenticated sessions. Instead of re-logging in on every run, you create a profile, log in once inside a browser session, and the cookies persist server-side.

# Create a profile
curl -X POST "https://api.browser-use.com/api/v2/profiles" \
  -H "X-Browser-Use-API-Key: $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"name": "LinkedIn Account"}'
 
# List your profiles
curl "https://api.browser-use.com/api/v2/profiles" \
  -H "X-Browser-Use-API-Key: $API_KEY"

You can also import your existing browser cookies using the Browser Use Chrome extension — which means any site you're already logged into locally can be used immediately without re-authentication.

Pricing

Sessions are billed per minute with unused time refunded. New accounts get $10 free credit — roughly 165 hours of session time at PAYG rates, or 500 tasks at task pricing.

Browser Use vs. Alternatives

How to Install

clawdbot install browser-use

Then configure your API key (get one at cloud.browser-use.com — new signups receive $10 free credit):

clawdbot config set skills.entries.browser-use.apiKey "bu_your_key_here"

Practical Tips

1. Create profiles before your first automation run. Set up a profile, spin up a session, log into all the sites you'll need to access, then stop the session. That profile will hold your authenticated state for subsequent runs.

2. Use task subagents for data extraction, sessions for interactive workflows. If you need to pull a price, extract a table, or fill a form with known inputs — use tasks. If you need Claude to make decisions as it navigates (e.g., handle dynamic content, respond to unexpected states) — use a session.

3. Set maxSteps conservatively on new tasks. Browser automation can loop unexpectedly. Start with maxSteps: 20 when testing a new task and increase once you've confirmed the agent is navigating correctly.

4. Stop sessions when done — immediately. Unused time is refunded, but you have to stop the session explicitly. If Claude finishes a task, issue the stop command right away rather than letting the timeout handle it.

5. Use flashMode for speed-sensitive tasks. When running high-volume extraction tasks where the sites are well-structured, flashMode significantly reduces time-per-task at similar accuracy.

6. Use the Chrome extension to seed profiles. Rather than logging into accounts inside a cloud browser (which requires managing 2FA, captchas, etc.), use the Browser Use Chrome extension to sync cookies directly from your local Chrome profile. Much faster for accounts with complex auth flows.

Considerations

• Cost accumulates with long sessions: At $0.06/hour, an idle 4-hour session costs $0.24. Not expensive, but if you're running many concurrent sessions, monitor your usage through the Browser Use dashboard.
• Site anti-automation policies: Browser Use includes anti-detection measures, but aggressive sites (Google, LinkedIn, Twitter) actively combat automation. Account-level bans are possible for high-volume usage.
• Task accuracy varies by site complexity: Simple structured sites (e-commerce product pages, form submissions) work reliably. Sites with heavy JavaScript, CAPTCHAs, or unusual UI patterns may require vision: true or manual session control.
• External API dependency: Unlike local Playwright, this skill requires network access to api.browser-use.com. Downtime or API changes affect all users.
• Cost of failed tasks: Failed tasks are still billed for the steps executed. Test new tasks with small maxSteps limits to minimize wasted spend.

The Bigger Picture

browser-use represents the cloud-native future of browser automation: no local setup, no infrastructure maintenance, and no CSS selectors. By abstracting the browser into an API with persistent profiles and a purpose-built LLM, it makes authenticated web automation accessible to anyone who can write a natural language prompt.

For AI agents that need to operate across the web — monitoring competitors, extracting structured data, automating form submissions, or managing SaaS dashboards — this is the lowest-friction path from "I want to automate X on Y website" to a working implementation.

View the skill on ClawHub: browser-use