MCP Camoufox
Bypasses Cloudflare's anti-bot detection systems to access websites protected by Cloudflare security measures.
Provides browser automation capabilities using a stealth-modified Firefox fork with C++ level anti-detection features.
Enables automated login and interaction with GitHub while avoiding bot detection and blocking.
Allows automated login and interaction with Google services while bypassing anti-bot detection systems.
Provides image hosting integration through the README's use of Imgur for hosting the project logo.
Enables JavaScript execution within browser pages for dynamic content interaction and automation.
Runs as a Node.js application with version 18+ requirement for server execution.
Distributed as an npm package with installation via npx for easy deployment.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@MCP Camoufox scrape job listings from LinkedIn for software engineer roles in San Francisco"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
MCP Camoufox
An MCP server with 135 tools that gives an AI agent a Camoufox browser and an HTTP client with Firefox's TLS fingerprint.
Camoufox is a Firefox build that spoofs its fingerprint in C++ rather than through injected JavaScript. This server drives it with Playwright over Juggler, Firefox's automation protocol. Requests that need no browser go through impit, which presents a Firefox TLS fingerprint.
Use the browser when a page needs JavaScript, a login or a person looking at it. Use HTTP when you only need the data.
✅ Camoufox with its fingerprint spoofing, driven over Juggler instead of CDP
✅ HTTP with a Firefox TLS fingerprint, reusing the browser's cookies when you want
✅ Replies that say what happened: a click reports whether the page changed, navigated or opened a tab
✅ Passwords, cookies and tokens hidden from tool output by default
✅ Logins kept in a persistent profile, or a temporary profile for one-off sessions
✅ Search through an endpoint you control (SearXNG, Brave, Tavily, Exa), with no result-page scraping
✅ Works with Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Droid and other MCP clients
Install
You need Node.js 18 or newer. The Camoufox browser downloads itself the first time browser_launch is called: about 1.3 GB, usually 2 to 5 minutes, about 2.4 GB on disk (measured on macOS arm64 with Camoufox 158). The download runs in the background. browser_launch answers within a minute with how far it got and asks to be called again, and server_status shows the progress. Tools that need no browser, such as scrape_markdown and http_request, work meanwhile. A browser deleted to free disk space is downloaded again the same way.
Add the server to your client.
Claude Code
All projects:
claude mcp add camoufox --scope user -- npx -y mcp-camoufox@latestCurrent project only:
claude mcp add camoufox -- npx -y mcp-camoufox@latestCodex CLI
Add to ~/.codex/config.toml:
[mcp_servers.camoufox]
command = "npx"
args = ["-y", "mcp-camoufox@latest"]OpenCode
Add to ~/.config/opencode/opencode.json:
{
"mcp": {
"camoufox": {
"type": "local",
"command": ["npx", "-y", "mcp-camoufox@latest"],
"enabled": true
}
}
}Other clients
Open the one you use:
Edit the config file, then restart the app:
macOS:
~/Library/Application Support/Claude/claude_desktop_config.jsonWindows:
%APPDATA%\Claude\claude_desktop_config.jsonLinux:
~/.config/Claude/claude_desktop_config.json
{
"mcpServers": {
"camoufox": {
"command": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}Use ~/.cursor/mcp.json for all projects, or .cursor/mcp.json in one project:
{
"mcpServers": {
"camoufox": {
"command": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}Use ~/.windsurf/mcp.json for all projects, or .windsurf/mcp.json in one project:
{
"servers": {
"camoufox": {
"command": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}Use your VS Code settings or ~/.continue/config.json for all projects, or .vscode/mcp.json in one project:
{
"mcpServers": {
"camoufox": {
"command": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}Edit ~/.config/zed/settings.json:
{
"context_servers": {
"camoufox": {
"command": {
"path": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}
}Add to ~/.gemini/settings.json:
{
"mcpServers": {
"camoufox": {
"command": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}Add to ~/.factory/mcp.json:
{
"mcpServers": {
"camoufox": {
"type": "stdio",
"command": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}Add to ~/.gemini/antigravity/mcp_config.json:
{
"mcpServers": {
"camoufox": {
"command": "npx",
"args": ["-y", "mcp-camoufox@latest"]
}
}
}Add to ~/.trae/mcp.json. Here mcpServers is an array:
{
"mcpServers": [
{ "name": "camoufox", "command": ["npx", "-y", "mcp-camoufox@latest"] }
]
}On Windows the browser goes to %LOCALAPPDATA%\camoufox\camoufox\Cache and the data to %USERPROFILE%\.camoufox-mcp\. Output paths accept C:\path\file.json and C:/path/file.json.
Related MCP server: CloakMCP
Usage examples
An agent usually reads the page, then acts on the refs it got back:
browser_launch(url="https://example.com")
browser_snapshot()
click(ref="e3")
fill(ref="e5", value="hello")Read a page without starting a browser. smart_fetch only uses the browser if the response looks blocked and one is running:
scrape_markdown(url="https://example.com/docs")
smart_fetch(url="https://shop.example.com/item/42")Log in with the browser, then call the site's API over HTTP with the same cookies:
browser_launch(url="https://app.example.com/login", headless=false)
smart_fill(fields_json='{"Email":"me@example.com","Password":"..."}', submit_label="Sign in")
wait_for_change()
http_request(url="https://app.example.com/api/orders")Find out why a click did nothing, before spending the click:
assert_clickable(selector="#submit")
# FAIL: covered by div.overlayRun a sequence in one call. A failed step stops it and tells you where to resume:
workflow_run(steps=[
{tool: "navigate", args: {url: "https://example.com"}},
{tool: "smart_fill", args: {fields_json: '{"Search":"camoufox"}'}},
{tool: "click_text", args: {text: "Search"}},
{tool: "wait_for_change", args: {}}
])Search through an endpoint you control:
search(query="firefox tls fingerprint", endpoint="http://127.0.0.1:8899")Test results
These come from the screenshots below, taken when each was captured. Sites and detectors change, so read them as what was observed and not as a guarantee. A pass on a test page does not mean a given protected site will let you in: IP reputation and session history count too.
Site | What it checks | Result |
| Fingerprint tests | ✅ Passes, apart from the expected "Chrome missing" |
| WebDriver, User-Agent, CDP, Navigator | ✅ All four "Normal" |
| Turnstile widget, with Cloudflare's test key | ✅ Widget reaches "Success!" |
| Full-page "Just a moment..." challenge | ⚠️ Playwright click rejected, 0 of 3. Real OS cursor click passed, 2 of 2, on macOS from one machine and IP |
Measured against Playwright MCP
Both servers were driven as real MCP clients over stdio on one machine (Apple M4 Pro, macOS), headless, with a new server and browser per round. mcp-camoufox ran Camoufox 158.0.1-beta.41. The other server was Microsoft's Playwright MCP 0.0.85 with its own Firefox 159. Numbers are medians of 5 rounds for speed and 3 for actions. The machine was in use while they ran, so treat small gaps as noise.
Measure | mcp-camoufox | Playwright MCP |
Start the browser and load a first page | 1.6 to 2.1 s | 1.0 to 1.1 s |
Navigate again to a local page | 59 ms | 60 ms |
Navigate again to a Wikipedia article | 219 ms | 1,091 ms |
Click that changes the page | 52 ms | 551 ms |
Click on a disabled button | refused in 431 ms | timeout after 5,015 ms |
Click on a covered button | 2,574 ms, clicked through the fallback, blocker named | timeout after 5,018 ms |
Click that opens an | 429 ms, reported as the dialog | 55 ms |
Click and fill inside a cross-origin iframe | 71 ms and 25 ms | 561 ms and 27 ms |
| 0.6 ms | 508 ms |
Memory with one form page open | 1,324 MB | 1,049 MB |
Snapshot of a 158-element form | 1,866 tokens | 3,963 tokens |
Snapshot of a long Wikipedia article | 20,229 tokens, cut at 60,000 characters and says so | 235,191 tokens |
Tool definitions sent to the client | 14,956 tokens for 135 tools | 4,568 tokens (it has 25) |
DOM calls a page can observe during a click | 0 | 1 |
| false | true |
| Normal | Robot |
| 30 of 57 rows pass, 1 fails ("Chrome missing") | 27 of 57 pass, 4 fail |
Where it loses: the browser takes half a second to a second longer to start, and nearly all of that is the Camoufox binary itself. It uses more memory unless you launch with ublock=false (then about 1,040 MB). Its tool list costs a client more than three times as many tokens. A click that changes nothing waits a full second before saying so, where Playwright MCP returns after half a second without saying whether anything happened.
The two snapshots are not the same thing. mcp-camoufox lists interactive elements and headings. Playwright MCP returns the whole accessibility tree including text. Token counts use the o200k_base tokenizer as an estimate.
Cloudflare Turnstile
The widget shows "For testing only": this demo uses Cloudflare's test sitekey and returns a dummy token. It shows that the click lands on the widget. It does not show that a protected site lets you through.
click_turnstile reports what it observed after the click: whether a token appeared, the final title and URL, whether a challenge still shows, and whether the sitekey is a Cloudflare test key. A click that produced no token is an error. A cf_clearance cookie on its own is not proof of passing either.
A Playwright click carries mozInputSource=0 and an empty pointerType, and a real challenge can reject that. turnstile_coords followed by os_click_xy delivers a real OS cursor click instead, as two separate calls, in a headed browser. macOS is tested, with its Accessibility permission. The Windows and Linux paths are written but not tested. It moves your actual cursor and brings the window forward, so it will interrupt what you are doing on that machine.
bot.sannysoft.com
navigator.webdriver is missing and WebDriver Advanced passes. The one red row is "Chrome missing", which is correct for a browser presenting itself as Firefox.
browserscan.net
WebDriver, User-Agent, CDP and Navigator all read "Normal".
Tools
135 tools in 32 groups. docs/TOOLS.md lists every tool and parameter, and a test fails if that list or this page stops matching the running server.
Group | Tools | For example |
Browser Lifecycle | 4 |
|
Navigation | 4 |
|
Tabs | 4 |
|
Viewport | 2 |
|
Scroll | 2 |
|
Frames | 2 |
|
Group | Tools | For example |
DOM & Content | 7 |
|
Element Inspection | 5 |
|
Accessibility | 1 |
|
Smart Selectors | 3 |
|
Scraping & Extraction | 4 |
|
Group | Tools | For example |
Element Interaction | 12 |
|
Keyboard | 2 |
|
Mouse XY | 6 |
|
Humanize / Anti-Bot | 5 |
|
Dialog | 2 |
|
Batch Operations | 6 |
|
Compound | 4 |
|
Workflow Helpers | 3 |
|
Wait | 7 |
|
JavaScript | 2 |
|
Assertions | 4 |
|
Group | Tools | For example |
Cookies | 3 |
|
Local Storage | 3 |
|
Session Storage | 3 |
|
Session Portability | 7 |
|
Storage Inspection | 4 |
|
Session Warmup & Detection | 2 |
|
Group | Tools | For example |
Console & Network | 9 |
|
Browserless HTTP | 5 |
|
Debug | 6 |
|
Site Automation | 2 |
|
Main features
A click reports what the page did: navigated, changed, opened a tab, or nothing within a second.
A click on a disabled or hidden element is refused. A left click blocked by a covering element falls back to synthetic pointer events, names what covered it, and is an error if nothing then changed.
A blocked right or double click is refused. It is never replayed as a left click.
click_textandclick_rolerefuse to guess between several matches and list them.A dialog that appears during any call is named in that call's reply, with how it was answered.
fill_formstops at a failed field and does not submit.batch_actionsstops at the first failure by default. A failedfillleaves the field's old value in place.browser_closechecks that the process is gone. On Windows it cannot check, and says so.A call that cannot finish answers within its deadline instead of hanging.
Cookie values, Authorization, Set-Cookie, passwords and tokens are masked by default in network_get, network_get_detail, wait_for_request, wait_for_response, intercept_log, export_curl, export_har, http_request, http_session_cookies, cookie_list, cookie_export, storage_diff, localstorage_get and sessionstorage_get. This covers tokens in JSON bodies, form bodies and URL parameters such as access_token. Pass reveal_values=true when you need the real value.
Password and other secret fields are masked in browser_snapshot, inspect_element, form_introspect, find_by_label, get_attribute, get_html, fill, type_text and error messages. reveal_values does not apply to these.
The masking goes by field name and known token shapes, so a secret in free text under an ordinary name can still get through. evaluate returns what your script returns. navigate, get_url and tab_list show the page's URL as it is. Session files from storage_state_save, auth_capture and cookie_export_file, and HAR files, are written with mode 0600.
Path | Contents |
| Default profile: cookies, localStorage, cache |
| Temporary profiles from |
| Screenshots and default HAR exports |
|
|
browser_launch(fresh_profile=true) uses a temporary profile, removed on browser_close. The server also removes it when the client disconnects, and the next launch removes one left by a crashed server. reset_profile deletes the default profile with its logins, so call it only when you mean that.
Session cookies die with the browser even in the persistent profile. Use cookie_set(expires_days=30) or storage_state_save to keep a login.
browser_launch asks for a spoofed screen that holds the window, and with headless=false one that holds your largest display too. If no fingerprint has a screen that large, it shrinks the viewport to fit and says so, because a viewport larger than its own screen is something a detector can read. fingerprint_audit prints what a page sees and flags contradictions. It checks consistency only. It cannot tell you whether a particular site will accept you.
The persistent profile keeps the same screen, GPU and core count across launches, where before ten launches of one profile showed ten different devices with the same logins; the canvas noise still changes per launch, and fresh_profile=true draws a new device every time.
With headless=false the page follows the window as you resize it, and the screen the page sees is as large as your largest display, so resizing does not create a fingerprint mismatch. Headless uses a fixed viewport; change it with set_viewport_size.
Limits
save_pdfcannot work here: Playwright generates PDFs only on Chromium. Usescreenshot(full_page=true).Firefox has no CDP, so there are no performance traces, coverage, heap snapshots or CPU and network throttling.
browser_snapshotlists elements inside visible iframes (refs likef1e3, up to 8 frames and 60 elements per frame) and inside open shadow roots (up to 1,000 roots, 8 levels deep), but cannot read a closed shadow root.A fingerprint that claims another operating system than the machine's own starts about 4 seconds slower and still lists the machine's speech voices.
os="auto"is the default for that reason. This was measured on macOS only.CLS is unavailable, because Firefox does not implement
layout-shift.Camoufox occasionally never commits a navigation. A 30-second timeout from
navigateis usually this, and a retry works.tab_newretries by itself.No CAPTCHA solving. Use
headless=falseand solve it yourself.npm auditmay flagadm-zipthroughcamoufox-js. npm applies overrides only from the root project, so add"overrides": { "adm-zip": "0.6.0" }to your ownpackage.jsonif your scanner requires it.
Troubleshooting
Problem | What to do |
"Browser not running" | Call |
"The browser is running but every tab is closed" | Call |
"A copy of Camoufox is already open" | Another browser is using the default profile. Close it, or launch with |
A click reports no page change | The page may react later, so try |
A ref stopped working | Refs belong to one snapshot. Call |
"Camoufox … is being downloaded now" | The first launch, or the browser was deleted. Wait a minute and call |
"Camoufox could not be downloaded" | The message says why. |
| You passed |
The first launch after an install fails with "Version information not found" | You are on 0.12.0 or older, which fetched the browser with a newer camoufox-js than it could read. Upgrade. |
| You are on a version before 0.9.3. Upgrade. |
A tool says it did not finish | The page may be wedged. |
I cannot see the browser window |
|
How it works
AI agent
| MCP over stdio
v
mcp-camoufox (Node.js) ---- impit (Firefox TLS) ---> HTTP requests with no browser
| Playwright over Juggler
v
Camoufox (patched Firefox)browser_snapshot returns a short list of the elements on the page, each with a ref, so an agent can act by ref instead of reading the whole page. The refs are kept in a registry in Camoufox's isolated JavaScript world. Nothing is written into the document, so a page's own scripts cannot find them.
The server's own reads (snapshot, click checks, change detection) also run in that isolated world. A test page that wraps querySelector, getBoundingClientRect and similar functions counted zero calls from a snapshot, a click, a fill and a hover. evaluate is the exception, because it runs your code: by default it runs beside the page's scripts so it can reach their variables, and the page can observe what that code calls. Pass world="isolated" when you only need the DOM.
Related
mcp-stealth-chrome is the sister project: the same idea for a real Chrome over CDP, in Python.
License
MIT
Available Tools
133 toolsaccessibility_snapshotA
Get accessibility tree snapshot — compact view of page structure for LLM understanding.
Inputs: no named parameters.
assert_clickableA
Hit-test an element WITHOUT clicking: would a real click actually land on it? Answers "why did my click do nothing?" before you spend the click. Reports the element that would intercept the click when something covers it.
Inputs: ref, selector.
assert_element_visibleA
Assert element exists and is visible. Returns success/fail (no throw).
Inputs: selector (required).
assert_text_presentA
Assert text is present anywhere on page (case-sensitive substring).
Inputs: text (required).
assert_url_matchesC
Assert current URL matches pattern (substring or regex).
Inputs: pattern (required), regex.
auth_captureA
Save current session as named auth state (e.g. logged-in user). Convenience wrapper: storage_state_save to ~/.camoufox-mcp/sessions/.json
Inputs: name (required).
back_and_snapshotB
Navigate back + return snapshot.
Inputs: no named parameters.
batch_actionsC
Execute multiple actions in one call. Each action: {type, ref?, value?, text?, key?, url?}.
Inputs: actions (required).
browser_closeA
Close the browser. Cookies are preserved in the persistent profile (~/.camoufox-mcp/profile). If the launch used fresh_profile=true, the temp profile is removed.
Inputs: no named parameters.
browser_launchA
Launch Camoufox stealth browser and navigate to URL. Browser persists between calls. By default cookies/localStorage persist in ~/.camoufox-mcp/profile. Set fresh_profile=true to start with a clean temp profile (auto-cleaned on browser_close) — useful when switching between accounts on the same domain.
Inputs: url, geoip, width, height, +5 more.
browser_recoverA
Escape hatch when the browser is wedged and browser_close can't complete: force-drops the connection and resets server state so browser_launch works again. Also reports a profile lock held by another Camoufox process, which is the usual cause of 'A copy of Camoufox is already open'.
Inputs: no named parameters.
browser_snapshotA
Get visible interactive elements with ref IDs. Use refs with click/fill. Always call after navigation. On large pages (Outlook, dashboards) the response can be truncated — narrow with roles=["button","textbox"] or paginate with offset/limit. Refs stay stable regardless of filters (every visible element is still numbered).
Inputs: limit, roles, offset.
chatgpt_generate_batchA
Generate MANY images on chatgpt.com IN PARALLEL (one tab per job, fire-all-then-collect) and save each to disk. Submits every job first WITHOUT waiting, then waits for all generations concurrently — far faster than sequential. For a CONSISTENT feed set, pass shared_image_paths (e.g. [logo] and/or a style-reference image like a previously-generated hero) uploaded to EVERY tab, plus style_suffix (a shared style spec) appended to every prompt. Requires an authenticated chatgpt.com session (import cookies first). Returns per-job results (saved path / ok / bytes / error).
Inputs: jobs (required), stagger_ms, timeout_ms, style_suffix, +1 more.
chatgpt_generate_imageA
Generate or edit an image on chatgpt.com end-to-end and save it to disk in ONE call. Opens a fresh chat, optionally uploads reference images (e.g. a brand logo), submits the prompt, waits for the generated image to finish, then writes the result PNG to output_path. Requires an authenticated chatgpt.com session (import cookies first via cookie_import). Returns the saved path and pixel dimensions. Note: chatgpt image generation is slow (~60-120s) — set timeout_ms accordingly.
Inputs: prompt (required), output_path (required), timeout_ms, image_paths.
checkC
Check checkbox or radio button.
Inputs: ref (required).
clickA
Click element by ref ID from browser_snapshot. Auto JS-fallback for overlays.
Inputs: ref (required), button, dblclick.
click_and_snapshotA
Click element by ref + wait + return snapshot. Perfect for buttons that trigger navigation/dialog.
Inputs: ref (required), wait_ms.
click_and_waitA
Click element then wait for navigation or selector. Atomic — fewer roundtrips than separate click + wait_for.
Inputs: ref, selector, timeout_ms, wait_for_url, +1 more.
click_at_cornerA
Click a corner of an element — close/X buttons, delete icons and dismiss controls usually live there, not in the centre.
Inputs: ref, corner, offset, selector.
click_element_offsetB
Click at a percentage position inside an element instead of its centre — for wide labels whose real checkbox sits at the left edge, sliders, or split buttons.
Inputs: ref, selector, x_percent, y_percent.
click_roleA
Click element by ARIA role and name. Same ambiguity guard as click_text: several matches → candidate list, not a guess.
Inputs: role (required), name, index, within.
click_textA
Click element by visible text. If the text matches several elements it FAILS with a numbered candidate list instead of silently clicking the first one — narrow with within ("@dialog", a CSS selector, or "ref:e5") or pick one with index.
Inputs: text (required), exact, index, within.
click_turnstileA
Auto-solve Cloudflare Interactive Turnstile checkbox. Locates the widget via in-page selectors AND the Playwright frame API (handles closed shadow roots that document.querySelector misses), polls for render, skips if already solved, then does a humanized real-mouse click with retries + small nudge, verifying the cf-turnstile-response token after each attempt. Managed Challenge full-page interstitials still need mcp-stealth-chrome.
Inputs: offset_x, offset_y, max_attempts, wait_render_ms.
console_getB
Get captured console messages.
Inputs: no named parameters.
console_startA
Start capturing console messages from all tabs.
Inputs: no named parameters.
cookie_deleteC
Delete cookies. Both empty = clear all.
Inputs: name, domain.
cookie_exportA
Export all cookies as JSON string. Use with cookie_import to transfer session.
Inputs: domain.
cookie_export_fileA
Export all cookies to a JSON file (Playwright format).
Inputs: path (required).
cookie_importA
Import cookies from JSON (from cookie_export). Restores session state.
Inputs: cookies_json (required).
cookie_import_fileB
Import cookies from a JSON file (Playwright format).
Inputs: path (required).
cookie_listC
List cookies.
Inputs: domain.
cookie_setA
Set a cookie. IMPORTANT: with expires_days=0 this creates a SESSION cookie, which Firefox keeps in memory only — it is gone after browser_close even though the profile itself persists. Pass expires_days (e.g. 30) to write a login session that survives a relaunch.
Inputs: name (required), value (required), domain (required), path, +4 more.
describe_pageA
Compact LLM-friendly page summary (title, heading, key buttons, forms). Cheaper than browser_snapshot for agent context. Also returns intent — a classified hint of what kind of page this is (login_email, login_password, otp_input, captcha, stay_signed_in, protect_account, error_page, consent, two_factor, account_disabled, content, unknown) — so the agent can branch with one read.
Inputs: no named parameters.
detect_anti_botA
Heuristic detection of anti-bot vendor on current page (Cloudflare, DataDome, Akamai, PerimeterX, Imperva).
Inputs: no named parameters.
detect_content_patternA
Auto-detect repeated content patterns (cards, listings, rows) and suggest CSS selectors. Run this BEFORE extract_structured to find the right selectors.
Inputs: min_items.
dialog_auto_handleA
Install a PERSISTENT dialog handler that stays armed across every dialog and every tab (dialog_handle is one-shot). Reads its action at fire time, so you can change it without re-arming. Set enabled=false to remove it.
Inputs: action, enabled, prompt_text.
dialog_handleA
Set handler for the next alert/confirm/prompt on ANY open tab (first dialog wins, handler then clears).
Inputs: action, prompt_text.
drag_and_dropB
Drag from one element to another.
Inputs: source_ref (required), target_ref (required).
evaluateC
Execute JavaScript in page context.
Inputs: expression (required).
export_curlA
Rebuild a captured request as a runnable curl command, so you can replay or share an API call without re-deriving its headers. Needs network_start(capture_bodies=true) — headers are not recorded otherwise.
Inputs: id, redact, url_contains.
export_harA
Export captured network traffic as a HAR 1.2 file (openable in DevTools/other HAR viewers). Requires network_start first; headers and bodies are only included when network_start(capture_bodies=true) was used.
Inputs: path, limit.
extract_structuredA
Extract structured data from repeated elements (cards, rows, listings). Auto-deduplicates, filters empty items, extracts direct text only. Use detect_content_pattern first to find correct selectors.
Inputs: fields (required), container_selector (required), limit, deduplicate_by, +1 more.
extract_tableC
Extract data from an HTML table as JSON array.
Inputs: limit, selector.
fillA
Fill input/textarea by ref ID. Always replaces existing content (email/number inputs are cleared explicitly first — Firefox's select-all is a no-op on those, which would otherwise append).
Inputs: ref (required), value (required).
fill_formC
Fill multiple form fields and optionally submit.
Inputs: fields (required), submit_ref.
find_by_labelA
Find input element by its label text (). Returns ref + how many matched.
Inputs: label (required), within.
find_by_placeholderB
Find input by placeholder text. Returns ref + how many matched.
Inputs: placeholder (required), within.
find_by_textA
Find elements by visible text — returns EVERY match (total + a ref and ancestor path per candidate), so you can tell whether the one you want is really the one you'd click. Skip browser_snapshot when you know the text.
Inputs: text (required), exact, limit, within.
fingerprint_auditA
Report the fingerprint a site actually sees from this browser, and flag internal contradictions — the mismatches that get an automated browser detected are far more revealing than any single value. Read-only: it inspects, it does not change anything.
Inputs: no named parameters.
form_introspectA
Analyse a form in one call: per field the label, type, current value, required/pattern/length constraints, validation state, and the JS framework it is bound to. Tells you what to fill and why a submit is being rejected without guessing from a snapshot.
Inputs: form_selector.
frame_evaluateB
Execute JavaScript inside a specific frame/iframe.
Inputs: expression (required), frame_name, frame_index.
get_attributeB
Get a specific attribute value from an element.
Inputs: ref (required), attribute (required).
get_htmlC
Get HTML content from page or element.
Inputs: outer, selector.
get_linksA
Get all links on the page with URL and text.
Inputs: filter.
get_page_errorsA
Get uncaught JavaScript errors + unhandled promise rejections from the current page. Captured by a hook installed at browser_launch, so the buffer resets on every navigation (read it before navigating away). Max 100 entries per page load.
Inputs: no named parameters.
get_textC
Get visible text from page or element.
Inputs: selector.
get_urlA
Get current URL and title.
Inputs: no named parameters.
get_viewport_sizeA
Get current viewport dimensions.
Inputs: no named parameters.
go_backB
Navigate back in history.
Inputs: no named parameters.
go_forwardB
Navigate forward in history.
Inputs: no named parameters.
hoverB
Hover over element by ref ID.
Inputs: ref (required).
http_requestA
HTTP request WITHOUT a browser, using a real Firefox TLS/HTTP2 fingerprint (impit). By default it reuses the live browser's cookies for that URL, so you can log in with the browser and then hit the site's API cheaply. Far faster and lighter than navigating — use it for APIs, JSON, and any page that doesn't need JS.
Inputs: url (required), body, proxy, method, +5 more.
http_session_cookiesA
Show which browser cookies would be sent with an HTTP request to this URL. Use it to confirm session sharing before relying on http_request.
Inputs: url (required).
humanize_clickA
Click element with humanized mouse approach (3-step Bezier-like curve before click). Use for anti-bot pages.
Inputs: ref, selector.
humanize_typeB
Type text with Gaussian-distributed delays between keystrokes (mean ~80ms, sigma ~30ms). Mimics human typing rhythm.
Inputs: text (required), ref, selector, mean_delay_ms.
indexeddb_deleteA
Delete an IndexedDB database by name for the current origin (clears SPA state that survives a cookie wipe).
Inputs: name (required).
indexeddb_listA
List IndexedDB databases for the current origin. Many SPAs keep auth state and drafts here, invisible to cookie/localStorage tools.
Inputs: no named parameters.
inject_init_scriptA
Inject a script that runs before every page load.
Inputs: script (required).
inspect_elementB
Get detailed info about an element (tag, attributes, bounding box, styles).
Inputs: ref (required).
intercept_logC
Show what interception blocked and allowed.
Inputs: limit, action.
intercept_startA
Block requests before they leave the browser. Blocking image/media/font/stylesheet typically removes most of a page's bytes — the single biggest speed-up for scraping — and block_urls kills trackers and ad frames. Routed on the context, so tabs opened later are covered too. Re-calling replaces the previous rules; intercept_stop removes them. 'document' cannot be blocked: it would abort the navigation itself.
Inputs: log_limit, block_urls, block_types.
intercept_stopA
Remove interception rules and report what they blocked.
Inputs: no named parameters.
list_framesB
List all frames/iframes in the page.
Inputs: no named parameters.
localstorage_clearA
Clear all localStorage.
Inputs: no named parameters.
localstorage_getB
Get all localStorage data or a specific key.
Inputs: key.
localstorage_setC
Set a localStorage item.
Inputs: key (required), value (required).
login_classicA
Composite login for classic email→password forms (Google, Microsoft, generic SSO). Auto-detects the email field, clicks Next/Continue on multi-step forms, fills the password, submits, and optionally enters a TOTP 2FA code. Collapses the usual 5–8 fill/click/snapshot calls into one. Heuristic — if a form is unusual, fall back to individual fill/click tools. Returns the step log + a fresh snapshot.
Inputs: email (required), password (required), totp_code, totp_secret, +2 more.
mouse_click_xyA
Click at exact x,y coordinates. steps>0 adds interpolated pre-movement (human-like).
Inputs: x (required), y (required), steps, button.
mouse_driftA
Random mouse movements over a duration — builds up mouse history before action (CF/DataDome behavior analysis).
Inputs: points, duration_ms.
mouse_moveB
Move mouse to x,y. steps>0 interpolates path (human-like).
Inputs: x (required), y (required), steps.
mouse_recordA
Start recording mouse positions in the page (replay with mouse_replay). Re-calling replaces any previous recorder instead of leaking its listener.
Inputs: max_points, duration_ms.
mouse_replayA
Replay last recorded mouse path with original timing.
Inputs: speed.
network_getA
List captured network requests in capture order, keeping the most recent limit rows. Each row shows an #id usable with network_get_detail.
Inputs: limit, filter.
network_get_detailA
Full request + response detail (headers and text body) for one captured request. Requires network_start(capture_bodies=true) BEFORE the request fired. Identify the request by id (from network_get) or by url substring.
Inputs: id, url.
network_startA
Start capturing network requests. With capture_bodies=true also records request/response headers + text bodies (json/text/xml/form only, capped at body_limit bytes) so you can inspect API payloads via network_get_detail — no need to pivot to evaluate()+fetch().
Inputs: body_limit, capture_bodies.
page_statsA
Page statistics: element count, size, load metrics. Use to decide extraction strategy.
Inputs: no named parameters.
paste_textA
Fill a field with a REAL paste. Puts the text on the clipboard and presses Ctrl/Cmd+V so the page receives a trusted paste event with actual clipboardData — the only thing that works for frameworks that listen to paste alone (Svelte 5 / SolidJS runes, some Qwik forms). Falls back to a synthetic event and then the native value setter.
Inputs: text (required), ref, selector.
performance_timelineA
Navigation + paint + resource timing for the current page: TTFB, DOMContentLoaded, load, FCP, LCP, and the 5 slowest resources. Note: Firefox does not implement layout-shift, so CLS is unavailable here.
Inputs: no named parameters.
press_keyB
Press key or combo (Enter, Escape, Control+a, etc.).
Inputs: key (required).
query_selector_allB
Query elements by CSS selector, return text/attributes of all matches.
Inputs: selector (required), limit, attribute.
reloadB
Reload the current page.
Inputs: no named parameters.
reload_and_snapshotB
Reload page + return snapshot.
Inputs: no named parameters.
reset_profileA
Delete the persistent profile (~/.camoufox-mcp/profile) entirely. Use to start fresh — cookies, localStorage, history all wiped. Browser must be closed first (call browser_close before this).
Inputs: no named parameters.
save_pdfA
Save page as PDF. NOTE: Playwright can only generate PDFs in headless Chromium — Camoufox is Firefox, so this fails by design here. Use screenshot(full_page=true) instead. Kept for API compatibility.
Inputs: path.
scrape_markdownA
Fetch a URL and return clean, LLM-ready markdown (headings, links, lists preserved; nav/footer/scripts stripped). Default path is browserless (impit) — fast and cheap. Set use_browser=true for JS-rendered pages (needs browser_launch).
Inputs: url (required), proxy, max_chars, impersonate, +1 more.
scrape_pageB
Smart page scraper — auto-detect and extract main content, links, metadata. Strips nav/footer noise.
Inputs: include_meta, include_links, max_text_length, only_main_content.
screenshotA
Screenshot the page, or a single element with ref/selector (great for documenting one modal). Returns the IMAGE inline — no second Read call — plus the saved path. Set return_image=false for path only.
Inputs: ref, name, selector, full_page, +1 more.
scrollC
Scroll the page.
Inputs: amount, direction.
scroll_toA
Scroll a specific element into view (the page-level scroll tool only moves by pixels).
Inputs: ref, block, selector.
searchA
Search the web through a search API YOU control — a self-hosted SearXNG, or Brave/Tavily/Exa with a key — and get back a normalised title/url/snippet list. This never scrapes a SERP: that shipped in 0.9.0 and was removed in 0.9.2 because engines answer some queries with confidently wrong results. A real API contract does not have that failure mode. Set MCP_SEARCH_ENDPOINT (and MCP_SEARCH_API_KEY when the provider needs one) to avoid passing them every call.
Inputs: query (required), count, api_key, endpoint, +3 more.
search_pageB
Grep the CURRENT page's visible text and return matches with surrounding context. Costs nothing compared to a snapshot or a screenshot — use it first to find where a term actually appears.
Inputs: text (required), regex, max_matches, context_chars, +1 more.
select_optionC
Select option from dropdown.
Inputs: ref (required), value (required).
server_statusB
Health check — verify server, browser status, active tabs.
Inputs: no named parameters.
sessionstorage_clearA
Clear all sessionStorage for the current origin (parity with localstorage_clear).
Inputs: no named parameters.
sessionstorage_getB
Get all sessionStorage data or a specific key.
Inputs: key.
sessionstorage_setC
Set a sessionStorage item.
Inputs: key (required), value (required).
session_warmupA
Visit innocuous public sites (Google, Wikipedia) to build browsing history before targeting protected site. Helps with CF/DataDome IP scoring.
Inputs: sites, duration_ms.
set_viewport_sizeA
Set the viewport width and height. This is the exact control over what the page renders and what a screenshot captures. It does NOT resize the OS window: Camoufox reports a frozen outerWidth/outerHeight, so asking for a viewport larger than the launch window makes the content bigger than its own window — impossible geometry that a detector can read. The reply says so when that happens.
Inputs: width (required), height (required).
smart_fetchB
Dual-mode fetch: tries the browserless HTTP path first and escalates to the stealth browser ONLY when the response looks anti-bot blocked. This is the efficiency core — high-volume reading stays cheap, the browser fires only when it's actually needed.
Inputs: url (required), proxy, max_chars, impersonate, +1 more.
smart_fillA
Fill form fields by their LABEL text (fuzzy, case-insensitive) instead of refs — no snapshot needed. Values go through the same clearing logic as fill, so email/number fields replace rather than append.
Inputs: fields_json (required), within, submit_label.
storage_diffA
Compare current storage against an earlier storage_snapshot — shows exactly which cookies/localStorage/sessionStorage keys were added, removed or changed. The fastest way to find which key holds a session token.
Inputs: name.
storage_snapshotA
Capture cookies + localStorage + sessionStorage + URL into a named slot, so storage_diff can tell you exactly what an action changed.
Inputs: name.
storage_state_loadA
Load cookies + localStorage from a JSON file (created by storage_state_save). Bypass CF/login if session is fresh.
Inputs: path (required), navigate_to.
storage_state_saveA
Save cookies + localStorage to a JSON file. Reload via storage_state_load on a fresh browser to skip login/CF entirely.
Inputs: path (required).
tab_closeA
Close a tab by index (-1 = active), or by url_contains.
Inputs: index, url_contains.
tab_listA
List all open tabs.
Inputs: no named parameters.
tab_newB
Open new tab.
Inputs: url.
tab_selectA
Switch to a tab by index, or by url_contains (first tab whose URL contains the substring).
Inputs: index, url_contains.
type_textA
Type text char by char via keyboard.
Inputs: text (required), delay.
uncheckC
Uncheck a checkbox.
Inputs: ref (required).
upload_fileC
Upload file to file input.
Inputs: ref (required), file_path (required).
wait_and_snapshotA
Wait for selector/text then return snapshot. Combines wait_for + browser_snapshot in one call.
Inputs: text, state, timeout, selector.
wait_forB
Wait for element/text to appear or disappear.
Inputs: text, state, timeout, selector.
wait_for_any_ofA
Race multiple wait conditions — returns the first that matches, so the agent can branch immediately without sequential probes. Each condition is {kind: 'selector'|'text'|'url_contains'|'title_contains', value: string}. Returns the index + kind + value of the winning condition (or 'timeout' if none matched). Ideal for post-login flows where the next page could be any of several (e.g. 'Stay signed in?', 'Skip for now', or the inbox directly).
Inputs: conditions (required), timeout.
wait_for_changeA
Wait until the page actually CHANGES and report what changed (url / title / DOM size / text). This is the honest version of a fixed sleep after a click: it returns as soon as something happened, or tells you nothing did.
Inputs: poll_ms, timeout, settle_ms.
wait_for_network_idleA
Wait until there are no in-flight requests for idle_ms continuously. Requests are counted from the moment the page was opened, so one that was ALREADY in flight when you call this is not mistaken for idle.
Inputs: idle_ms, timeout_ms.
wait_for_requestA
Block until the page ISSUES a request matching a URL substring (wait_for_response waits for the reply). Use it to confirm an action actually fired its API call.
Inputs: url_pattern (required), method, timeout.
wait_for_responseB
Wait for a network response matching a URL pattern.
Inputs: url_pattern (required), timeout.
wait_for_urlC
Wait for URL to match a pattern.
Inputs: pattern (required), timeout.
workflow_runA
Run a list of tool calls in sequence — any tool this server exposes, by name — and return a per-step log. Resumable: a failed run tells you the index, and start_at skips the steps that already succeeded. Each step is {"tool":"navigate","args":{"url":"…"},"label":"optional"}.
Inputs: steps (required), start_at, stop_on_error.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 133 tools
There are many near-synonymous tools for core actions: at least nine click variants, seven wait variants, and multiple snapshot/scrape/extract tools. The descriptions are detailed, but an agent still faces frequent ambiguity about which variant fits a given situation.
All names are snake_case and readable, but the verb/noun pattern is inconsistent: browser_launch and tab_new are noun_verb, get_url and cookie_list are verb_noun, and many tools are bare verbs like navigate, reload, hover, and scroll. There is no single predictable convention, though the mixed style is still navigable.
133 tools is an extreme count for any server, even a broad browser-automation toolkit. The surface area is bloated with specialized variants that could plausibly be consolidated, making the set hard to scan and reason about.
The domain is thoroughly covered: browser lifecycle, navigation, interaction, extraction, network capture, storage, cookies, dialogs, frames, anti-bot evasion, session management, and compound workflows all have dedicated tools. There are no obvious dead ends or major missing operations.
Maintenance
Related MCP Connectors
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
Web scraping MCP server — scrape, extract structured data, screenshot any site with anti-bot bypass.
Undetectable cloud browser sessions for AI agents and scrapers. Navigate, extract, click, captcha.
Stealth web browser for agents: search, fetch, click, download and type in persistent MCP sessions.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceAn MCP server for stealth browser automation that uses human-like interaction patterns to bypass bot detection via the Chrome DevTools Protocol. It enables users to navigate, interact with elements, and capture data from websites using undetectable behaviors like Bezier mouse movements and Gaussian typing delays.117 npm1MIT
- AlicenseAqualityCmaintenanceAn MCP server that provides LLMs with stealth browser automation capabilities via CloakBrowser to bypass bot detection services like Cloudflare and reCAPTCHA. It supports full page interaction, content extraction, and human-like behavior through 30 specialized tools.2052 PyPI14Apache 2.0
- -licenseNot gradedqualityNot gradedmaintenanceUndetectable browser automation server for MCP-compatible AI agents, offering 225 tools across 32 sections to navigate, extract, clone pages, and bypass antibot systems like Cloudflare.1-
- AlicenseAqualityDmaintenanceA stealth browser automation MCP server that wraps CloakBrowser's patched Chromium to bypass bot detection, providing 22 tools for web navigation, interaction, and session management.2456 PyPI2MIT