Give an agent a browser, and everything it saw along the way.

• Auto-waiting on actionability — fewer explicit waits in agent code
• Chromium out of the box, with Firefox and WebKit available to the engine
• Native request routing, which the network capture layer builds on
• The engine to reach for unless a site actively fights automation

• puppeteer-extra-plugin-stealth strips the usual automation fingerprints
• Direct CDP access when a tool call needs to go lower than the abstraction
• Leaner startup and memory footprint per session
• The engine for bot-protected targets

Levenshtein distance over visible text, so a typo or a re-worded label still resolves.

browser_find_element_by_text({
  tabId,
  text: "Sbumit Order",
  threshold: 0.8,
  elementType: "button"
})
// → matches "Submit Order"

Handles both HTML patterns — explicit <label for> and an input wrapped inside its label — plus aria-label and placeholder.

browser_find_input_by_label({
  tabId,
  labelText: "Email Address"
})
// → the input, however it was wired up

When nothing connects a field to its caption but position on screen, search outward from an anchor.

browser_find_element_near({
  tabId,
  anchorXpath: "//span[text()='Username']",
  direction: "below",
  maxDistance: 100
})

A parser, not a model call — it reads element type, text filter, visibility and state out of the phrase.

browser_query_dom({
  tabId,
  query: "visible clickable elements with delete text"
})
// → tag:button|a + text:delete + visible

Pair it with a screenshot: the agent sees a region, then asks what is actually inside those coordinates.

browser_find_elements_in_area({
  tabId,
  x: 840, y: 120,
  width: 360, height: 480
})

browser_create_session({ engine: "playwright" })
browser_launch_tab({ sessionId, url: "https://app.example.com/login" })

browser_find_input_by_label({ tabId, labelText: "Email" })
browser_input_text({ tabId, xpath: "//input[@id='email']", text: "agent@co.com" })

browser_click_element({ tabId, xpath: "//button[@type='submit']" })
browser_wait_for_element({ tabId, xpath: "//div[@data-view='dashboard']" })
browser_take_snapshot({ tabId, fullPage: true })

Start a recording and keep working. Every MCP tool call the agent makes lands in the transcript — navigations, clicks, typing, key presses, scrolls and waits.

browser_start_recording({ tabId })
// ... the agent does its work ...
browser_stop_recording({ recordingId })

Show the agent how it is done. Record yourself doing the flow by hand, upload it, and the agent replays your steps instead of rediscovering them.

• Point the extension at your server and project key
• Start, click through the flow, stop, upload
• The agent finds it with browser_list_user_recordings

Console output is collected per tab from the moment it opens — no flag to set, nothing to start. That includes uncaught exceptions, which is usually the line that actually explains the failure.

browser_get_console_logs({ tabId, level: "error", limit: 50 })

An agent that works out a page’s quirk can write it down and read it back later, tagged and searchable, instead of solving the same puzzle on every run.

browser_add_note({
  sessionId,
  content: "Cookie banner intercepts the first click. "
           "Dismiss #cc-accept before anything else.",
  tags: ["quirk", "shop.example.com"]
})

browser_list_notes({ sessionId, tag: "quirk" })

find and query number what they return, so you act on !1 rather than copying an XPath out of a table.

b4m open https://app.example.com   # session + tab, remembered
b4m find "Sign in"                 # numbered matches
b4m click !1
b4m type !2 "hunter2"
b4m shot                            # prints the screenshot URL
b4m logs -l error
b4m shell https://app.example.com  # interactive

-o json emits records rather than a rendered table, and colour disappears when stdout is not a terminal — so redirecting gives you a file, not escape codes.

b4m query links --json | jq -r '.[].attributes.href'
b4m run checkout-flow.json          # replay a Recorder file

# CI cannot complete a device flow — use a project key
export BROWSE4ME_API_KEY=proj_...
b4m open https://staging.example.com

• Clicks, text input, scrolls and navigation, as you perform them
• Written out as Chrome Recorder JSON — the same format the agent records
• Uploaded to your project, scoped to your API key

• Load the unpacked extension, set server URL and project key in options
• Hit Start recording in the popup and use the site normally
• Stop, then Upload
• The agent lists and replays it over MCP

Seven guides the agent can read without leaving the session — plus live resources for current sessions, tabs and recordings.

welcome://message
docs://best-practices
docs://search-strategies
docs://framework-clicks
docs://network-capture
docs://console-logs
docs://recording-guide