Browser automation using browser MCP tools for web scraping, form filling, testing, and security auditing. Controls a real Chromium through a 10-tool gateway that reaches ~98 underlying tools — navigation, element interaction, DOM querying, tab management, network monitoring, and JavaScript execution. Use when asked to browse, scrape, fill a form, automate a website, check a page, open a browser, take a screenshot, monitor network requests, run a security audit, bypass a Cloudflare challenge, or interact with any web page programmatically. Features a token-optimized DOM walker that reduces HTML to compact JSON with numeric IDs, plus Cloudflare Turnstile solving and fingerprint (user-agent/locale/timezone/geo) overrides.
Instrucciones de origen · Vista previa de solo lectura
name
browser-automation
description
Browser automation using browser MCP tools for web scraping, form filling, testing, and security auditing. Controls a real Chromium through a 10-tool gateway that reaches ~98 underlying tools — navigation, element interaction, DOM querying, tab management, network monitoring, and JavaScript execution. Use when asked to browse, scrape, fill a form, automate a website, check a page, open a browser, take a screenshot, monitor network requests, run a security audit, bypass a Cloudflare challenge, or interact with any web page programmatically. Features a token-optimized DOM walker that reduces HTML to compact JSON with numeric IDs, plus Cloudflare Turnstile solving and fingerprint (user-agent/locale/timezone/geo) overrides.
Use when the user wants to browse a website, scrape web content, fill out forms, automate browser interactions, take screenshots, run security audits, or interact with any web page. Examples: 'open this URL', 'scrape that page', 'fill out the form', 'take a screenshot', 'check this website', 'automate the login', 'run a security audit', 'monitor network requests'.
argument-hint
[url or task description]
context
fork
metadata
{"author":"coffeegrind123","version":"1.4"}
Browser Automation
The server runs in GATEWAY mode — only 10 tools are directly callable
Every other capability — screenshots, forms, cookies, localStorage, tabs,
network/console logs, JS execution, scrolling, stealth/user-agent, geolocation,
security audit, Cloudflare — exists on the server but is hidden to keep the
context small. Reach it in one hop:
search_tools(query="read cookies") → ranked 'name(params) — summary' lines
describe_tool(name="set_cookie") → full schema, when the summary is not enough
call_tool(name="set_cookie", arguments={...}) → actually run it
call_tool validates arguments against the real schema, so a mistyped field
returns a clear error rather than failing silently. Pass {} for a no-arg tool.
Throughout this document, any tool NOT in the ten above is written by its bare
name for readability — invoke it as call_tool(name="<that name>", arguments={…}).
If you are unsure a tool exists, search_tools is cheaper than guessing.
No tool works until start_browser() completes. No page tools work until
navigate() completes. Always stop_browser() when finished.
navigate() settles the page before it returns — do not add a wait after it.
It waits for the network to go quiet and reports how that ended, e.g.
Navigated to https://x/ (network idle after 0.6s, 246 requests). A read
straight afterwards sees rendered content. network still active after 10.0s is
normal for a page that polls or streams and usually still reads fine. Pass
settle=0 to return the instant navigation commits, when you will not read the
page.
Scenario
Call
Default (headful)
start_browser()
Headless scraping
start_browser(headless=true)
With proxy
start_browser(proxy="socks5://host:port")
Authenticated proxy
start_browser(proxy="http://user:pass@host:port")
Persist session
start_browser(user_data_dir="/path/to/profile")
Check if running
get_browser_status()
Primary Discovery: get_interaction_tree()
ALWAYS use get_interaction_tree() as the FIRST step to understand a page.
This is the single most important tool. It returns a token-optimized JSON
of all interactive elements — 96% smaller than raw HTML. Each element gets
a numeric ID you can click/type into directly.
t = type (btn, link, in, sel, chk, rad, tab, mnu, el)
l = label (inferred from aria-label/placeholder/text/title/alt)
r = region (hdr, nav, main, side, ftr, dlg)
v = current value (inputs only)
Use numeric IDs directly:
click(selector="1") — clicks element with id 1
type_text(text="hello", selector="2") — types into element with id 2
Read reference/dom-walker.md if you need the full type/region code reference or label inference rules.
Tool Selection Decision Tree
"I need to find something on the page"
Interactive element (button, link, input)?
YES → get_interaction_tree()
Only buttons? → find_buttons(filter_text="optional")
Only inputs? → find_inputs(filter_type="text|email|...")
NO → Visible text content?
YES → get_text_content()
NO → get_content() [raw HTML — use sparingly, large tokens]
"I need to interact with something"
Click → get_interaction_tree() → click(selector="<id>")
→ OR click(text="Button Text") if you know exact text
Fill form → fill_form(form_data="{\"#email\": \"a@b.com\", \"#pass\": \"123\"}")
NOTE: form_data is a JSON *string*, not an object. Serialize it.
→ OR type_text(text="value", selector="<id>") [core tool, direct]
Dropdown → select_option(selector="<css>", value="option_value")
Upload → upload_file(selector="input[type=file]", file_path="/path")
Submit → submit_form() or press_enter()
"I need to wait for something"
Page load → nothing. navigate() already settled it.
Element appear → wait_for_element(selector="css", timeout=10, visible=true)
Network idle → wait_for_network(timeout=5) # AFTER a click, not after navigate
Specific request → wait_for_request(url_pattern="api/data")
All three real waits are hidden tools — call_tool(name="wait_for_element", …).
That extra hop is exactly why you should not spend one on a page load that
navigate() has already handled.
A blind wait(seconds=N) is the last resort: slower than settling on a fast
page, too short on a slow one. Reach for it only when the thing you are waiting
for is neither an element nor a request — an animation, or a rate limit.
"I need to extract data"
DOM data → execute_js("document.querySelector('.price').textContent")
Visible text → get_text_content()
API response → clear_logs() → [action] → wait_for_request(url_pattern="api/")
Screenshot → screenshot(save_path="/tmp/page.png")
"The page is blocked by Cloudflare"
Probe → is_cloudflare_challenge_present() # fast, no clicking
Solve → bypass_cloudflare(timeout=20) # clicks the Turnstile checkbox
Most Cloudflare-gated sites pass automatically when the browser is headed
(started with start_browser(headless=false)); under --headless=new the
Turnstile is often unsolvable. For an interactive challenge that survives headed
mode, bypass_cloudflare() solves it (wraps zendriver's built-in verify_cf).
set_user_agent / set_locale / set_timezone / set_geolocation align the
fingerprint with a proxy's geo when needed.
Do not pass low_memory=true when stealth matters: its flags (software WebGL,
--disable-gpu) are themselves a bot signal.
When bypass_cloudflare() times out — check for a hard-fail before retrying
bypass_cloudflare() polls for a Turnstile checkbox. If Turnstile has already
fingerprinted the browser and rejected it, it never renders one — it renders a
feedback report instead, and the call can only ever time out. Retrying, raising
the timeout, or re-clicking cannot fix this. Confirm with get_content():
.../auto/failure (and a visible cf-turnstile-feedback wrapper) = fingerprint
rejection. Stop clicking and change approach.
Try a crawler User-Agent first when the goal is public SEO/markup. Many sites
allowlist search crawlers by UA string alone, with no reverse-DNS verification, so
plain curl walks straight past the challenge and returns the real HTML — no
browser, no solver, far faster and more reliable:
Worked on a self-hosted Turnstile gate (2026-07) that hard-failed the automated
browser but served robots.txt / sitemap.xml / full HTML to a Googlebot UA.
Note a self-hosted widget (its own <form action="/verify_captcha"> with a
data-sitekey) is NOT the standard CF edge challenge — the edge-challenge advice
above may not apply to it at all.
Reach for this only for publicly published content (markup, robots, sitemaps)
that the site already serves to crawlers by design — not to get at anything gated,
paid, or private.
General Rules
Autonomy: Execute the full browsing workflow without asking for
confirmation at each step. Ask only when genuinely ambiguous.
Token discipline: Prefer get_interaction_tree() over get_content().
Raw HTML burns context fast. Only use get_content() for HTML structure
the interaction tree doesn't expose.
Do not wait after navigate: navigate() settles the page itself and says
how it ended. Add a wait only after a click or an action that starts new
traffic — wait_for_element() for content an SPA hydrates in late, or
wait_for_network() after a submit.
3a. An empty read tells you which problem you have.get_text_content() on a
page with no text returns [chars 0-0 of 0] followed by the reason:
still loading (read again, or navigate with a bigger settle), no body element, has markup (the text is in an iframe or shadow root — try
get_content or get_interaction_tree), or really is blank. Read that line
before concluding the site blocked you.
Numeric IDs are ephemeral: They change on every get_interaction_tree()
call. Get the tree, use IDs immediately. Never cache across navigations.
CSS selectors for stability: When referencing the same element across
page changes, use a CSS selector instead of a numeric ID.
execute_js syntax: Scripts must NOT start with return. Simple
expressions directly. Complex logic: (function(){ var x = 1; return x; })()
Network monitoring: Call clear_logs() BEFORE the action you want to
monitor, then get_network_logs() or wait_for_request() AFTER.
Tab hygiene: Close tabs opened with new_tab() when done.
Error recovery: If a click or navigation fails, screenshot() to
visually diagnose. The page may have changed or a modal may block.
Auth persistence: Use get_cookies() / set_cookies() for cookie-based auth.
For JWT-based SPAs, use get_local_storage() / set_local_storage() — many
modern apps store tokens in localStorage, not cookies.
Common Workflows
Core tools are called directly; everything else goes through call_tool.
Web Scraping
start_browser(headless=true) → navigate(url) # settled on return
→ get_text_content() # or get_interaction_tree()
→ call_tool(name="execute_js", arguments={"script": "..."})
→ stop_browser()
Form Filling
start_browser() → navigate(url)
→ get_interaction_tree()
→ call_tool(name="fill_form", arguments={"form_data": "{\"#email\": \"user@ex.com\", \"#password\": \"secret\"}"})
→ click(selector="<submit_id>")
→ call_tool(name="wait_for_network", arguments={"timeout": 5}) # the click, not the navigate
→ call_tool(name="screenshot", arguments={"save_path": "/tmp/after.png"})
→ stop_browser()
It names the fix: still loading / no body / markup-but-no-text / blank
Unknown tool name
It is hidden, not absent
search_tools(query="…") then call_tool
Click does nothing
Element not visible
scroll_to_element() then retry
Element not found
Page not loaded yet
wait_for_element() then re-query
Stale numeric ID
Page mutated
Re-call get_interaction_tree()
execute_js fails
Started with return
Remove return, use IIFE
Network logs empty
Didn't clear first
clear_logs() before action
Form submit no effect
Need explicit submit
submit_form() or press_enter()
Interaction tree empty
SPA hydrates after the network settles
wait_for_element() on something you expect, or re-navigate with a larger settle; find_inputs()/find_buttons() as a fallback
Clicks blocked by overlay
Cookie banner or modal
Dismiss banner first: click(text="Accept All")
Critical Pitfalls
❌ Do NOT call any tool before start_browser() — everything will fail
❌ Do NOT use get_content() as primary discovery — wastes tokens massively
❌ Do NOT cache numeric IDs across navigations — they regenerate each call
❌ Do NOT start execute_js scripts with return — use bare expression or IIFE
❌ Do NOT add a wait() after navigate() — it already settled the page, and
under the gateway that wasted wait costs a whole call_tool hop
❌ Do NOT leave tabs open — close with close_tab() when done
❌ Do NOT read network logs without clearing first — stale entries mislead
❌ Do NOT use mouse_click(x,y) unless no other option — coordinates are fragile
❌ Do NOT call get_interaction_tree() then ignore the IDs — that's the whole point
❌ Do NOT interact with elements before dismissing cookie consent banners — they block clicks
❌ Do NOT assume auth is in cookies — check get_local_storage() for JWT tokens too
❌ Do NOT pass timeout to screenshot — it has no such parameter (only
save_path and full_resolution), and call_tool validates arguments, so it
is a hard error. Screenshots really can hang on an unresponsive page; guard
against that by settling the page first (which navigate now does) rather than
by inventing an argument
Self-Refinement Protocol
After completing a browser automation task, if you discovered something
non-obvious (a selector pattern, a timing requirement, a site-specific
quirk), append it to LEARNINGS.md:
## YYYY-MM-DD: <brief title>
- **Context**: What you were doing
- **Learning**: What was non-obvious
- **Rule**: The new rule to follow
Read reference/tools.md if you need full parameter documentation for any tool.
Read reference/patterns.md if you need automation recipes beyond the common workflows above.
Read reference/dom-walker.md if the interaction tree output is unclear or you need type/region code details.
