Skip to main content
Glama

🧭 TabPilot


Most browser MCP servers launch a fresh, blank browser instance. That works for scraping static pages, but fails completely on tasks that matter:

  • 🚫 Cloudflare & Bot Shields immediately flag fresh automation browsers.

  • 🚫 Corporate SSO, Okta, & 2FA make authenticating from scratch painful or impossible.

  • 🚫 Read-only tab viewers can only look at DOM text, not click or fill forms.

TabPilot bridges this gap. It attaches directly to your existing, logged-in Google Chrome. Your AI agents (Claude, Cursor, Antigravity, Cline) can read pages with up to 99% token savings, click elements with authentic mouse events (isTrusted: true), defeat React state caching traps, solve multi-row survey grids, and capture background screenshots without stealing window focus.

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚  YOUR WORKSPACE                     TABPILOT MCP ENGINE           REAL BROWSER    β”‚
β”‚                                                                                   β”‚
β”‚  β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”          β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”      β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚
β”‚  β”‚ AI Agents            β”‚          β”‚ TabPilot MCP Server  β”‚      β”‚ Real Chrome  β”‚ β”‚
β”‚  β”‚ - Claude Desktop     β”‚  stdio / β”‚ - Token Budget Slicerβ”‚ CDP  β”‚ - Active SSO β”‚ β”‚
β”‚  β”‚ - Cursor IDE         β”‚ ───────> β”‚ - Native Form Setters│────> β”‚ - Cookies    β”‚ β”‚
β”‚  β”‚ - Antigravity/Gemini β”‚   SSH    β”‚ - Matrix Grid Driver β”‚<──── β”‚ - 2FA Saved  β”‚ β”‚
β”‚  β”‚ - Cline / Windsurf   β”‚          β”‚ - Dual CDP/AppleScr. β”‚      β”‚ - Real Fingerβ”‚ β”‚
β”‚  β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜          β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜      β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

⚑ Quick Start

1. Test your environment in one command

Run tabpilot doctor via uvx (no installation required):

uvx tabpilot-mcp doctor

(Or install via pip: pip install tabpilot-mcp)

2. Configure your MCP Client

Add to claude_desktop_config.json or .cursor/mcp.json:

{
  "mcpServers": {
    "tabpilot": {
      "command": "uvx",
      "args": ["tabpilot-mcp"]
    }
  }
}

3. Launch Chrome with Remote Debugging

Quit Chrome completely, then start it pointing to your persistent automation profile:

# macOS
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
  --remote-debugging-port=9222 --user-data-dir="$HOME/tabpilot-chrome"

# Linux
google-chrome --remote-debugging-port=9222 --user-data-dir="$HOME/tabpilot-chrome"

# Windows
taskkill /F /IM chrome.exe
"C:\Program Files\Google\Chrome\Application\chrome.exe" ^
  --remote-debugging-port=9222 --user-data-dir="%USERPROFILE%\tabpilot-chrome"
IMPORTANT

🍎 macOS Zero-Config Fallback:
On macOS, if you do not launch Chrome with debugging flags, TabPilot automatically falls back to AppleScript to drive the Chrome you already have open.
Prerequisite: In Google Chrome, go to View β†’ Developer β†’ check "Allow JavaScript from Apple Events".


Related MCP server: Browser MCP

πŸ’° How TabPilot Saves 90%+ Tokens

Naive browser automation dumps document.body.innerText or raw HTML, wasting 15,000–45,000 tokens on navigation bars, cookies banners, and tracking scripts.

TabPilot provides Progressive Token Slicing:

Raw Page Dump (eval_js body)  β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ 18,000+ tokens
read_tab() (Link-Density)     β–ˆβ–ˆβ–ˆβ–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘ ~1,200 tokens (93% saved)
read_tab(selector)            β–ˆβ–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘ ~350 tokens (98% saved)
query_dom(selector)           ▏                                      ~18 tokens (99.9% saved)

Method

Tokens

Savings

Best For

query_dom(selector)

~15–30

99.9%

Checking button states (disabled), inputs, badges, or alerts

read_tab(selector)

~300–600

98.0%

Reading isolated articles, ticket cards, or form sections

read_tab() (Link-Density)

~1,200–1,800

93.0%

Full-page reads with nav, ads, headers, and footers stripped

❌ Naive eval_js innerText

15,000–35,000

0%

Context budget incinerator

πŸ’‘ Enforced Character Budget: Reads default to --max-chars 20000. If truncated, TabPilot alerts the model with exact cut sizes, prompting it to narrow down with selector instead of hallucinating.


πŸ₯Š Driving Forms That Fight Back

Modern Single-Page Applications (React, Vue, Svelte) defeat standard automation scripts. TabPilot solves the 3 most infamous SPA traps:

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ TRAP 1: React Synthetic Cache                                                     β”‚
β”‚ Standard `.value = 'x'` is ignored by React's internal tracker.                   β”‚
β”‚ βœ… TabPilot calls native property setters + fires both `input` and `change`.      β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚ TRAP 2: Multi-Row Survey Grid Batching                                            β”‚
β”‚ Clicking 25 matrix rows in one JS tick causes React to commit only the last row.  β”‚
β”‚ βœ… `fill_matrix` clicks row-by-row across microtasks with delays, then confirms.  β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚ TRAP 3: Virtual Portals (React-Select, Headless UI)                               β”‚
β”‚ No `<select>` exists; options render only after clicking the trigger.             β”‚
β”‚ βœ… `select_option_ui` clicks trigger, awaits portal mount, and selects option.    β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

βš–οΈ Comparison

Capability

Read-Only Tab Readers

Headless Bots (Puppeteer/Playwright)

🧭 TabPilot

Uses Existing Sessions & Cookies

βœ…

❌ Fresh empty profile

βœ… Real Chrome (No re-login)

Bypasses Cloudflare & Bot Shields

βœ… Human

❌ Bot fingerprint

βœ… Human browser fingerprint

Form Driving & Clicking

❌ Read-only

βœ…

βœ… Full Bidirectional Driving

React Synthetic Event Fix

❌

⚠️ Often missed

βœ… Native prototype setters

Survey Matrix Handler

❌

❌ Batched drops

βœ… fill_matrix task scheduling

Token-Optimized Extraction

⚠️ Basic text

❌ Raw DOM / costly vision

βœ… query_dom (18 tokens)

Background Tab Screenshots

❌ Must be active

βœ…

βœ… Off-screen CDP captures

Headless Ubuntu 24/7 Daemon

❌ macOS only

⚠️ Complex Docker

βœ… Native systemd stack

External Dependencies

Minimal

❌ Heavy Node/Playwright binaries

βœ… Zero dependencies beyond MCP


🐧 24/7 Headless Ubuntu Server Stack

TabPilot is engineered to run permanently on cloud VPS servers (AWS, Hetzner, DigitalOcean) with zero exposed ports:

[ Laptop / Client ] ──( Encrypted SSH Pipe )──> [ Remote Ubuntu Server ]
                                                        β”‚
                                                        β–Ό
                                             [ TabPilot MCP Server ]
                                                        β”‚ (CDP 127.0.0.1:9222)
                                                        β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ Managed Systemd User Stack (Restart=always, Linger Enabled)                       β”‚
β”‚   tabpilot-xvfb.service   ──>   tabpilot-wm.service   ──>   tabpilot-chrome.service
β”‚   (Virtual Framebuffer)         (Openbox Window Mgr)        (Real Chrome + SSO)   β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
  1. One-Command Setup: bash deploy/ubuntu/install.sh configures Xvfb, Openbox, Chrome, and systemd services with Restart=always.

  2. Font Protection (Anti-Tofu): Headless servers often render Vietnamese and CJK characters as empty square boxes. install.sh installs Noto and Liberation fonts; doctor validates text rendering.

  3. Cgroup Memory Caps: Prevent Chrome from exhausting host RAM:

    tabpilot install-stack --memory-high 4G
  4. Zero Remote Attack Surface: Chrome CDP stays locked to 127.0.0.1. Remote MCP connections run securely through SSH pipes (ssh you@server "tabpilot serve").


🧰 Tools & Resources Reference

TabPilot exposes 19 tools and 2 live resources:

πŸ›‘οΈ CAPTCHA Detection & Solving

  • detect_captcha(url_pattern, tab_id): Structured, zero-side-effect detection of CAPTCHAs (Cloudflare Turnstile/Interstitial, reCAPTCHA v2/Enterprise, hCaptcha, image grids, text captchas, and slider puzzles) with confidence scores, candidate IDs, and available strategies.

  • solve_captcha(tab_id, url_pattern, operation, candidate_id, solve_id, observation_id, action_id, action, strategy, agent_vision, timeout_ms, max_attempts, max_rounds, expected, activate_on_fail): Multi-step stateful solver supporting automated checkbox solvers, passive wait, and interactive Agent Vision observation-action loops (select_tile, drag, type_answer, verify) with postcondition access verification and token freshness checks.

πŸ” Reading & DOM Inspection

  • query_dom(selector, attrs, limit, visible_only): Atomic element inspection (~18 tokens). Checks disabled, checked, values.

  • read_tab(selector, mode, max_chars, url_pattern): Link-density markdown extraction with token caps.

  • list_tabs(url_pattern): Lists open tabs with IDs, titles, and URLs (regex filterable).

  • eval_js(expression, timeout_ms): Runs arbitrary JS in tab context; automatically awaits Promises in CDP.

🧭 Navigation & Tabs

  • open_tab(url, activate, wait_for_load): Opens a URL in a new tab, awaiting page completion.

  • close_tab(url_pattern, tab_id): Closes tab and polls until process confirms destruction.

  • navigate(url, wait_for_load): Points tab to a new URL and awaits document load.

  • activate_tab(url_pattern, tab_id): Brings target tab to foreground focus.

πŸ–±οΈ Interaction & Complex Forms

  • click(selector, text, nth): Scrolls element into view and emits trusted mouse event (isTrusted: true).

  • fill(selector, value, clear, press_enter): Sets form inputs via native prototype setters to trigger React/Vue.

  • select_option(selector, values, by): Selects options in native <select> or Select2 dropdowns.

  • select_option_ui(control_selector, option_text): Clicks trigger, waits for popup portal, and selects item.

  • wait_for(selector, state, text, timeout_ms): Polls until condition holds (visible, hidden, text, enabled).

πŸ“‹ Matrix Surveys & Evidence

  • scan_matrix(selector): Discovers multi-row grid questions and flags unanswered rows.

  • fill_matrix(column_index, rows, delay_ms): Answers matrix rows sequentially with task delays.

  • screenshot(full_page, selector, label): Offscreen capture of any tab (local path or inline base64).

  • browser_status(): Reports active backend, capabilities, CDP endpoint, and open tab count.

πŸ“¦ Live MCP Resources

  • tab://active β€” Markdown stream of the currently focused tab.

  • tab://{tab_id} β€” Markdown stream of any specific tab by ID.


πŸ› οΈ CLI Reference

tabpilot doctor                # Run diagnostic suite (Chrome, CDP, fonts, permissions)
tabpilot tabs                  # List open browser tabs in terminal
tabpilot serve                 # Start MCP server on stdio

# Linux Managed Stack Commands
tabpilot install-stack         # Install systemd user services
tabpilot up / tabpilot down    # Start or stop the headless stack
tabpilot status                # Check systemd stack status
tabpilot logs chrome -f        # Follow live Chrome logs

πŸ§ͺ Testing

git clone https://github.com/hunglp97/tabpilot-mcp.git
cd tabpilot-mcp
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Unit tests (Zero browser needed)
pytest -v

# Live integration tests (against Chrome on port 9222)
pytest -v -m live

πŸ“„ License

Distributed under the MIT License. See LICENSE for details.


Available Tools

17 tools
activate_tabA

Bring a target browser tab to the front and focus its window.

Side effects: Changes system window focus and switches the user's active tab viewport. Does not reload the page or alter DOM state.

Usage guidelines:

  • When to use: Use when a human user needs to observe the active page, or before capturing desktop-wide OS screenshots and video screencasts.

  • When NOT to use: Do NOT call this before reading or interacting with tabs. TabPilot tools (read_tab, query_dom, click, fill, eval_js, screenshot) work off-screen in background tabs without stealing focus.

Args: tab_id: Exact tab identifier (e.g. from list_tabs). If omitted, uses the frontmost tab or matches by url_pattern. url_pattern: Optional regex pattern matched against tab URLs (e.g. 'github.com').

Returns: Confirmation message containing the activated tab ID and title/URL. Returns an error message if no matching tab is found.

Inputs: tab_id, url_pattern. Returns: result

browser_statusA

Report which backend is connected, what it can do, and how many tabs are open.

Inputs: no named parameters. Returns: result

clickA

Click an element, scrolling it into view first.

On the CDP backend this is a real mouse event at the element's centre, so widgets that only react to genuine mousedown behave correctly.

Args: selector: CSS selector. text: Visible text to match. Combined with selector it narrows within those matches; alone it searches clickable elements. Exact matches win over partial ones, and the innermost match wins. tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. nth: Which match to click, 0-based.

Inputs: nth, text, tab_id, selector, +1 more. Returns: result

close_tabA

Close a tab.

Args: tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. Must identify one tab.

Inputs: tab_id, url_pattern. Returns: result

eval_jsA

Run a JavaScript expression in a tab and return its value as JSON.

The expression's value is returned, so write document.title, not return document.title. Promises are awaited on the CDP backend.

Do not use this to dump document.body.innerText β€” read_tab and query_dom answer those questions for a fraction of the tokens.

Args: expression: A JavaScript expression. tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. timeout_ms: Per-call timeout.

Inputs: expression (required), tab_id, timeout_ms, url_pattern. Returns: result

fillA

Set a form field's value so React, Vue and jQuery all notice.

Works on text inputs, textareas, contenteditable, checkboxes and radios (pass 'true'/'false' for those).

Args: selector: CSS selector for the field. value: Value to set. tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. clear: Empty the field first. nth: Which match to fill, 0-based. press_enter: Press Enter afterwards, for search boxes.

Inputs: value (required), selector (required), nth, clear, +3 more. Returns: result

fill_matrixA

Answer a matrix/grid question one row at a time, then verify the result.

Rows are clicked in separate JavaScript tasks with a delay between them. Clicking them in a loop instead makes React batch the updates and commit only the last row, leaving the rest blank and the page stuck failing validation.

Args: tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. selector: Override the matrix container selector. question_index: Which matrix question, as numbered by scan_matrix. column_index: Which answer column to pick. Negative counts from the right, so -1 is the last column. rows: Specific row indices. Defaults to whichever rows need answering. only_unanswered: Skip rows that already have an answer. delay_ms: Delay between rows. Raise it to 150-250 if rows stay blank.

Inputs: rows, tab_id, delay_ms, selector, +4 more. Returns: result

list_tabsA

List open tabs with their ids, titles and URLs.

Args: url_pattern: Optional regex; only tabs whose URL matches are listed.

Inputs: url_pattern. Returns: result

open_tabA

Open a URL in a new tab and return its id.

Args: url: Absolute URL to open. activate: Bring the new tab to the front. wait_for_load: Wait until the document finishes loading.

Inputs: url (required), activate, wait_for_load. Returns: result

query_domA

Inspect specific elements: tag, text, attributes, size, visibility.

Use this instead of reading the page when the question is about particular elements β€” form fields, buttons, error messages, option lists.

Args: selector: CSS selector. tab_id: Exact tab id. Prefer url_pattern. url_pattern: Regex matched against tab URLs. limit: Maximum elements to return. attrs: Only report these attributes. Defaults to a useful set (id, name, type, href, value, placeholder, role, aria-label, data-testid, class) plus disabled/checked/selected state. text_max: Truncate each element's text at this many characters. visible_only: Skip elements that are not rendered.

Inputs: selector (required), attrs, limit, tab_id, +3 more. Returns: result

read_tabA

Read a tab's content, converted to markdown and capped in size.

Args: tab_id: Exact tab id from list_tabs. Prefer url_pattern. url_pattern: Regex matched against tab URLs. mode: 'readable' (markdown, boilerplate stripped), 'text', or 'html'. selector: Read only this element's subtree. Use it β€” it is the cheapest way to keep a read small and on-topic. max_chars: Character budget. Defaults to the server's --max-chars. include_links: Keep links as markdown in 'readable' mode.

Inputs: mode, tab_id, selector, max_chars, +2 more. Returns: result

scan_matrixA

List the matrix/grid questions on a page and which rows are unanswered.

Call this before fill_matrix to see the shape of the question and to confirm the row indices.

Args: tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. selector: Override the matrix container selector.

Inputs: tab_id, selector, url_pattern. Returns: result

screenshotA

Capture a tab and save it to disk.

Captures the tab that was asked for even if it is not frontmost, and works headless β€” so this is also how evidence gets captured on a server. Needs the CDP backend.

Args: tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. full_page: Capture the whole scrollable page, not just the viewport. selector: Capture only this element. image_format: 'png' or 'jpeg'. quality: JPEG quality, 1-100. Ignored for PNG. label: Included in the filename, to make the file findable later. return_image: Inline the image in the result. Defaults on when the client is on another machine (TABPILOT_REMOTE / --return-images), since a saved path means nothing there.

Inputs: label, tab_id, quality, selector, +4 more.

select_optionA

Choose option(s) in a native or a Select2 widget.

Pass several values for a multi-select. If the element turns out to be a React-Select style widget with no underlying , this says so and you should use select_option_ui instead.

Args: selector: CSS selector for the . values: Option values or labels to select. tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. by: Match options by 'value', 'label', or 'auto' (either). nth: Which match to use, 0-based.

Inputs: values (required), selector (required), by, nth, +2 more. Returns: result

select_option_uiA

Choose an option in a React-Select / Headless UI dropdown by driving its UI.

These have no to set, so the menu is opened, awaited, and the option clicked β€” which cannot be done in a single JavaScript call because the menu does not exist yet when the click would fire.

Args: control_selector: CSS selector for the control that opens the menu. option_text: Visible text of the option to choose. tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. option_selector: Override the option selector if the defaults miss. timeout_ms: How long to wait for the menu to appear.

Inputs: option_text (required), control_selector (required), tab_id, timeout_ms, +2 more. Returns: result

wait_forA

Poll until a condition holds, then return. Fails with what it last saw.

Args: selector: CSS selector to watch. state: 'visible', 'hidden', 'present', 'absent', 'enabled', or 'text' (with the text argument). text: Substring to look for when state is 'text'. predicate: A JavaScript expression to poll instead of a selector, e.g. "document.readyState === 'complete'". tab_id: Exact tab id. url_pattern: Regex matched against tab URLs. timeout_ms: How long to keep polling. poll_ms: Interval between checks, minimum 50.

Inputs: text, state, tab_id, poll_ms, +4 more. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 17 tools

    Disambiguation5/5

    Each tool targets a distinct operation: reading page content (read_tab) vs inspecting specific elements (query_dom) vs running custom JS (eval_js), and interaction tools (click, fill, select_option, select_option_ui) have clear boundaries. Even similar tools like scan_matrix/fill_matrix and select_option/select_option_ui are explicitly distinguished by widget type and purpose.

    Naming Consistency4/5

    Most names follow a verb_noun pattern (list_tabs, read_tab, open_tab, activate_tab, fill_matrix), but a few deviate: navigate, click, fill, and screenshot are single verbs or nouns used as verbs, while browser_status and wait_for use different structures. The pattern is still predictable and readable.

    Tool Count4/5

    With 17 tools, the set is slightly larger than the ideal 3–15 range but each tool serves a concrete browser automation function (tab management, reading, interaction, waiting, screenshots, matrix handling). The count feels justified for the domain, not bloated.

    Completeness4/5

    The tool surface covers tab lifecycle (open, close, list, navigate), multiple read modes (markdown, DOM inspection, screenshots), form interaction (fill, click, select, matrix-specific tools), and synchronization (wait_for). Minor gaps exist like no explicit back/forward or keyboard shortcuts, but these can be worked around with navigate or eval_js.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    • Instapaper MCP server for ChatGPT, Claude and any AI agent: list unread articles, save URL, archive article, move to folder, read highlights and fetch article text in your own Instapaper account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to Instapaper yourself; Cotlet does not store your password and you need no Instapaper API key. About the official Instapaper API: Instapaper has a full REST API, API v2, and no official MCP server (checked 17 Oct 2026). Instapaper launched API v2 on 29 Sep 2026 with OAuth 2, personal access tokens, an OpenAPI spec and Python and TypeScript SDKs. Setup and FAQ: https://cotlet.dev/integrations/instapaper-api-mcp-server

    • Thumbtack MCP server for ChatGPT, Claude and any AI agent: list leads, reply to lead, send quote, update weekly budget, list reviews and export past leads in your own Thumbtack account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to Thumbtack yourself; Cotlet does not store your password and you need no Thumbtack API key. About the official Thumbtack API: Thumbtack has an official Partner Platform API for leads (called negotiations), messages and reviews, but its docs are gated and access is by request for integration partners (checked 11 Oct 2026). A pro gets no API key for their own account. Setup and FAQ: https://cotlet.dev/integrations/thumbtack-api-mcp-server

    • TaxDome MCP server for ChatGPT, Claude and any AI agent: create job, request documents, send invoice, message client and list overdue tasks in your own TaxDome account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to TaxDome yourself; Cotlet does not store your password and you need no TaxDome API key. About the official TaxDome API: TaxDome has no public API that firms can use today (checked 14 Oct 2026). Its Zapier integration only creates, updates and finds contacts, and TaxDome MCP, which would let Claude or ChatGPT work in a firm account, is announced but not open yet and is offered to Professional and Business plan firms first. Setup and FAQ: https://cotlet.dev/integrations/taxdome-api-mcp-server

      https://cotlet.dev/mcp/v1/taxdome
    • Loxo MCP server for ChatGPT, Claude and any AI agent: list jobs, read a job pipeline, search people, log candidate activity, move candidate to a stage, email a pipeline stage, read the activity report and create person in your own Loxo account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to Loxo yourself; Cotlet does not store your password and you need no Loxo API key. About the official Loxo API: Yes, Loxo has an Open API, but it is a paid feature: Loxo Support turns on the Settings > API Keys page for your account, and only account admins can use it (checked 22 Oct 2026). Loxo has no official MCP server, and its API reference has no endpoints for reports or for sending email. Setup and FAQ: https://cotlet.dev/integrations/loxo-api-mcp-server

      https://cotlet.dev/mcp/v1/loxo

    Related MCP Servers

    • A
      license
      A
      quality
      A
      maintenance
      Enables AI agents to drive your real, logged-in Chrome browser with existing sessions and cookies, bypassing CAPTCHA and anti-bot measures, with support for multi-session and human-in-the-loop workflows.
      40
      64
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI assistants to drive the user's real, already-logged-in Chrome browser locally, so they can read and navigate pages, click, type, fill forms, take screenshots, and inspect cookies, tabs, console, and network traffic without headless browsers or remote-debugging ports.
      1 npm
      ISC
    • A
      license
      Not graded
      quality
      A
      maintenance
      Lets AI coding tools drive the real, already-logged-in Chrome browser over local stdioβ€”snapshotting and reading pages, clicking, typing, managing tabs, capturing screenshots, and inspecting network, console, storage, and cookiesβ€”without opening a remote-debugging port. It installs as a Chrome extension plus native-messaging bridge, so agents work in the user's authenticated sessions while passing bot detection.
      1 npm
      ISC