browsirai

name	browsirai
description	Control your live Chrome browser session via MCP tools. Navigate, click, fill forms, take screenshots, read accessibility trees, inspect source code locations, and automate browser interactions.

browsirai

MCP server that connects AI coding agents to a running Chrome browser via Chrome DevTools Protocol. Interact with your live session -- logged-in state, cookies, and all open tabs.

Prerequisites

• A Chromium-based browser (Chrome, Edge, Brave, Arc) with remote debugging enabled
• Node.js 22+

Enable Remote Debugging

Option A -- Launch Chrome with a flag:

# macOS
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222

# Linux
google-chrome --remote-debugging-port=9222

# Windows
chrome.exe --remote-debugging-port=9222

Option B -- Enable in browser: navigate to chrome://inspect/#remote-debugging and toggle the switch.

Supported Browsers

Chrome, Edge, Brave, Arc, Vivaldi, Opera, Chromium. Discovery works on macOS, Linux, and Windows.

Quick Start

# Connect to browser on default port (9222)
browser_connect

# Or specify a custom port/host
browser_connect { "port": 9222, "host": "127.0.0.1" }

# List all open tabs
browser_tabs

# Navigate to a URL
browser_navigate { "url": "https://example.com" }

# Take a snapshot to see the page structure
browser_snapshot

Core Workflow

The snapshot-ref interaction pattern is the primary way to work with page elements:

1. Snapshot -- Call browser_snapshot to get the accessibility tree. Each element receives an @eN ref (e.g., @e1, @e2, @e15).
2. Interact -- Use the @eN ref in browser_click, browser_fill_form, browser_hover, etc.
3. Verify -- Call browser_screenshot to visually confirm the result, or browser_snapshot again to check updated state.

Example snapshot output:

@e1 heading "Welcome" level=1
@e2 link "Sign In"
@e3 textbox "Email" value=""
@e4 textbox "Password" value=""
@e5 button "Log In"
@e6 checkbox "Remember me"

After a snapshot, refs remain valid until the page navigates or the DOM changes significantly. If a tool fails with "invalid ref", take a new snapshot.

Tools Reference

Navigation

browser_navigate

Navigate to a URL. Waits for the page to load.

browser_navigate { "url": "https://github.com" }
browser_navigate { "url": "https://example.com", "waitUntil": "networkidle" }

Param	Type	Default	Description
url	string	required	URL to navigate to
waitUntil	"load" | "domcontentloaded" | "networkidle"	"load"	When to consider navigation complete

browser_navigate_back

Navigate back or forward in browser history.

browser_navigate_back
browser_navigate_back { "direction": "forward" }

Param	Type	Default	Description
direction	"back" | "forward"	"back"	History navigation direction

browser_wait_for

Wait for a condition before proceeding. Default timeout: 30 seconds.

browser_wait_for { "text": "Successfully saved" }
browser_wait_for { "textGone": "Loading..." }
browser_wait_for { "selector": ".results-table" }
browser_wait_for { "selector": ".modal", "state": "hidden" }
browser_wait_for { "url": "**/dashboard**" }
browser_wait_for { "fn": "document.querySelectorAll('.item').length > 5" }
browser_wait_for { "time": 2 }

Param	Type	Description
text	string	Wait until text appears in page body
textGone	string	Wait until text disappears from page body
selector	string	Wait until a CSS selector matches an element
state	"hidden"	Combined with selector -- wait until element is hidden
url	string	Wait until URL matches glob pattern (supports * and **)
fn	string	Wait until JS expression evaluates to truthy
time	number	Simple delay in seconds
timeout	number	Override timeout (seconds if <=60, milliseconds if >60)

Observation

browser_snapshot

Capture the accessibility tree with @eN refs. This is the primary tool for understanding page structure.

browser_snapshot
browser_snapshot { "compact": true }
browser_snapshot { "interactive": true }
browser_snapshot { "selector": "#main-content" }
browser_snapshot { "depth": 3 }

Param	Type	Default	Description
selector	string	-	CSS selector to scope the snapshot
compact	boolean	false	Hide InlineTextBox nodes and empty wrappers
interactive	boolean	false	Only show interactive elements (buttons, links, inputs, etc.)
cursor	boolean	false	Include elements with cursor:pointer style
depth	number	100	Maximum tree depth

Interactive roles included when interactive: true: button, link, textbox, checkbox, radio, combobox, listbox, menuitem, menuitemcheckbox, menuitemradio, option, searchbox, slider, spinbutton, switch, tab, treeitem.

browser_screenshot

Take a screenshot. Returns base64-encoded image data.

browser_screenshot
browser_screenshot { "fullPage": true }
browser_screenshot { "selector": "#hero-section" }
browser_screenshot { "format": "jpeg", "quality": 80 }
browser_screenshot { "annotate": true }

Param	Type	Default	Description
selector	string	-	CSS selector to screenshot a specific element
fullPage	boolean	false	Capture full scrollable page, not just viewport
format	"png" | "jpeg"	"png"	Image format
quality	number	-	JPEG quality (0-100). Only for jpeg format
annotate	boolean	false	Overlay numbered labels on interactive elements

browser_html

Get raw HTML content of the page or a specific element.

browser_html
browser_html { "selector": "#app" }

Param	Type	Description
selector	string	CSS selector to scope HTML output

browser_tabs

List open browser tabs. Only shows page-type targets (excludes service workers, extensions).

browser_tabs
browser_tabs { "filter": "*github.com*" }

Param	Type	Description
filter	string	Glob-style URL filter pattern

browser_console_messages

Retrieve captured console messages.

browser_console_messages
browser_console_messages { "level": "error" }
browser_console_messages { "limit": 20 }

Param	Type	Description
limit	number	Maximum messages to return (default: 1000, most recent)
level	"log" | "info" | "warn" | "error"	Filter by minimum severity

browser_network_requests

List captured network requests.

browser_network_requests
browser_network_requests { "filter": "*api*" }
browser_network_requests { "includeHeaders": true, "includeStatic": false }
browser_network_requests { "limit": 10 }

Param	Type	Default	Description
filter	string	-	Glob-style URL filter
limit	number	-	Maximum requests to return (most recent)
includeHeaders	boolean	false	Include request/response headers
includeStatic	boolean	true	Include static resources (JS, CSS, images, fonts)

Interaction

browser_click

Click an element. Supports three targeting methods: ref, selector, or coordinates.

browser_click { "ref": "@e5" }
browser_click { "selector": "#submit-btn" }
browser_click { "x": 150, "y": 300 }
browser_click { "ref": "@e2", "newTab": true }

Param	Type	Description
ref	string	@eN ref from snapshot
selector	string	CSS selector
x, y	number	CSS pixel coordinates (see Coordinates section)
newTab	boolean	Open link in new tab (adds Meta/Ctrl modifier)

Must provide one of: ref, selector, or both x and y.

The click sequence: scrolls element into view, computes center coordinates from the box model, then dispatches mouseMoved -> mousePressed -> 50ms delay -> mouseReleased.

browser_fill_form

Fill a form field. Clears existing value before typing. Dispatches input and change events.

browser_fill_form { "ref": "@e3", "value": "[e-mail]" }
browser_fill_form { "selector": "#search-input", "value": "search query" }

Param	Type	Description
ref	string	@eN ref from snapshot
selector	string	CSS selector
value	string	Text to enter

Must provide either ref or selector.

Handles different field types automatically:

• textbox: focus -> clear -> insert text -> dispatch events
• checkbox/radio: clicks to toggle
• combobox (select): sets value directly
• slider: sets value and dispatches events

Note: Will not fill readonly or disabled fields (returns an error).

browser_type

Type text into the focused element or a specific ref. Unlike browser_fill_form, does NOT clear existing value first.

browser_type { "text": "Hello world" }
browser_type { "text": "search term", "ref": "@e3", "submit": true }
browser_type { "text": "slowly typed", "slowly": true }

Param	Type	Default	Description
text	string	required	Text to type
ref	string	-	@eN ref to focus before typing
slowly	boolean	false	Type character-by-character with key events
submit	boolean	false	Press Enter after typing

• Fast mode (default): uses Input.insertText -- single CDP call, instant.
• Slow mode (slowly: true): dispatches individual keyDown/char/keyUp per character. Use for apps that listen to individual key events.
• Use browser_type (not browser_evaluate) to enter text in cross-origin iframes.

browser_press_key

Press a keyboard key or key combination.

browser_press_key { "key": "Enter" }
browser_press_key { "key": "Tab" }
browser_press_key { "key": "Control+c" }
browser_press_key { "key": "Control+Shift+a" }
browser_press_key { "key": "Escape" }

Param	Type	Description
key	string	Key name or combination with + separator

Supported keys: Enter, Tab, Escape, Backspace, Delete, ArrowLeft/Up/Right/Down, Home, End, PageUp, PageDown, Space, F1-F12, plus any single character.

Modifier keys: Control, Shift, Alt, Meta. Combine with +: Control+c, Meta+a, Control+Shift+Tab.

browser_hover

Hover over an element. Triggers mouseover/mouseenter events.

browser_hover { "ref": "@e7" }

Param	Type	Description
ref	string	@eN ref from snapshot (required)

browser_drag

Drag from one element to another. Uses synthesized mouse events with intermediate move points.

browser_drag { "startRef": "@e3", "endRef": "@e8" }

Param	Type	Description
startRef	string	@eN ref for the drag source
endRef	string	@eN ref for the drop target

Drag sequence: mouseMoved(start) -> mousePressed(start) -> intermediate mouseMoved steps -> mouseMoved(end) -> mouseReleased(end).

browser_scroll

Scroll the page or a specific element.

browser_scroll { "direction": "down" }
browser_scroll { "direction": "down", "pixels": 500 }
browser_scroll { "selector": ".sidebar", "direction": "down" }
browser_scroll { "selector": "#target-element" }

Param	Type	Default	Description
direction	"up" | "down" | "left" | "right"	-	Scroll direction
pixels	number	300	Pixels to scroll
selector	string	-	Scrollable container, or element to scroll into view

Three modes:

• selector only (no direction): scrolls the element into view (centered)
• selector + direction: scrolls within that container
• direction only: scrolls the page viewport

browser_select_option

Select option(s) in a <select> element. Matches by value attribute or visible label text.

browser_select_option { "ref": "@e12", "values": ["us-east-1"] }
browser_select_option { "ref": "@e12", "values": ["Option A", "Option C"] }

Param	Type	Description
ref	string	@eN ref (required)
values	string[]	Values or label text to select

Dialog & File

browser_handle_dialog

Accept or dismiss JavaScript dialogs (alert, confirm, prompt, beforeunload).

browser_handle_dialog { "accept": true }
browser_handle_dialog { "accept": false }
browser_handle_dialog { "accept": true, "promptText": "my input" }

Param	Type	Description
accept	boolean	Accept (true) or dismiss (false) the dialog
promptText	string	Text to enter in a prompt dialog

Waits up to 5 seconds for a dialog to appear if none is currently pending.

browser_file_upload

Upload files to a file input element.

browser_file_upload { "ref": "@e9", "paths": ["/Users/me/photo.jpg"] }
browser_file_upload { "ref": "@e9", "paths": ["/tmp/doc1.pdf", "/tmp/doc2.pdf"] }

Param	Type	Description
ref	string	@eN ref for the file input (required)
paths	string[]	Absolute file paths to upload

Lifecycle

browser_connect

Connect to a running browser instance via CDP.

browser_connect
browser_connect { "port": 9222 }
browser_connect { "host": "192.168.1.10", "port": 9222 }

Param	Type	Default	Description
port	number	9222	Debug port
host	string	"127.0.0.1"	Host address

browser_list

List all discoverable browser instances (scans default ports 9222, 9229).

browser_list

Returns browser name, version, and WebSocket debugger URL for each instance found.

browser_close

Close or detach from browser tabs.

browser_close
browser_close { "force": true }
browser_close { "force": true, "targetId": "ABC123..." }
browser_close { "force": true, "closeAll": true }

Param	Type	Default	Description
force	boolean	false	Actually close tab(s) instead of just detaching
targetId	string	-	Specific tab to close
closeAll	boolean	false	Close all page tabs

Without force, simply detaches the session (tabs remain open).

browser_resize

Resize the browser viewport.

browser_resize { "width": 1280, "height": 720 }
browser_resize { "width": 375, "height": 812, "deviceScaleFactor": 3 }
browser_resize { "preset": "mobile" }

Param	Type	Description
width	number	Viewport width in CSS pixels
height	number	Viewport height in CSS pixels
deviceScaleFactor	number	Device pixel ratio (DPR) override
preset	string	Device preset name

Presets: mobile (375x667), tablet (768x1024), desktop (1280x720), fullhd (1920x1080), reset (clear override).

Must provide either width + height together, or a preset.

browser_resize { "preset": "reset" }

Use preset: "reset" to clear the device metrics override and restore the browser's native viewport.

Advanced Observation

browser_annotated_screenshot

Take a screenshot with numbered labels overlaid on interactive elements. Useful for identifying clickable elements visually.

browser_annotated_screenshot
browser_annotated_screenshot { "selector": "#main-content" }

Param	Type	Description
selector	string	CSS selector to scope the annotated area

Returns an image with numbered labels plus a text index mapping each label to its role, name, and @eN ref.

browser_inspect_source

Inspect a DOM element and return its source code location. Maps UI elements back to their source files -- file path, line number, and component name.

browser_inspect_source { "selector": "h1" }
browser_inspect_source { "ref": "@e5" }
browser_inspect_source { "selector": "nav" }

Param	Type	Description
ref	string	@eN ref from snapshot
selector	string	CSS selector

Must provide either ref or selector.

Returns:

• Element: tag name
• Component: nearest framework component name
• Source: file path, line number, column number
• Component Stack: full ancestor component chain with source locations

Supported frameworks:

• React: Walks the Fiber tree and parses jsxDEV() calls in Function.toString() to extract fileName/lineNumber embedded by @babel/plugin-transform-react-jsx-source (works with Vite, CRA, Next.js dev mode)
• Vue: Reads __vueParentComponent.type.__file
• Svelte: Reads __svelte_meta.loc

Note: Only works in development mode where source metadata is preserved. Production builds strip this information.

Code Execution

browser_evaluate

Evaluate JavaScript in the page context. Supports async expressions.

browser_evaluate { "expression": "document.title" }
browser_evaluate { "expression": "document.querySelectorAll('a').length" }
browser_evaluate { "expression": "await fetch('/api/status').then(r => r.json())" }

Param	Type	Description
expression	string	JavaScript expression to evaluate
frameId	string	Target frame ID for execution

Returns the evaluated result. Handles primitives, objects, DOM nodes (serialized to description), null, and undefined. Async expressions are awaited automatically.

Coordinates

Screenshot image dimensions differ from CSS pixel dimensions due to device pixel ratio (DPR).

Screenshot pixels = CSS pixels x DPR
CSS pixels = Screenshot pixels / DPR

• browser_screenshot detects DPR using a 3-level cascade: Page.getLayoutMetrics -> Emulation.getDeviceMetricsOverride -> window.devicePixelRatio
• browser_click coordinates (x, y) use CSS pixels
• Typical Retina display: DPR = 2, so divide screenshot pixel coordinates by 2

Security

Network requests are automatically redacted to prevent secret leakage:

• Headers: Authorization, Cookie, Set-Cookie, X-API-Key, X-Auth-Token, X-CSRF-Token, Proxy-Authorization, and other auth-related headers are replaced with [REDACTED]
• Body fields: password, secret, token, api_key, apiKey, access_token, refresh_token, client_secret, private_key are redacted in JSON bodies
• Inline secrets: JWTs (eyJ...) are replaced with [REDACTED_JWT], Bearer tokens with Bearer [REDACTED]

Tips & Best Practices

• Prefer browser_snapshot over browser_html for understanding page structure. The accessibility tree is semantic, compact, and provides @eN refs for interaction.
• Use @eN refs for reliable targeting. Refs are tied to accessibility tree nodes and resolve to exact DOM elements. CSS selectors can be fragile if the DOM changes.
• Use browser_type for cross-origin iframes -- browser_evaluate cannot access cross-origin content, but Input.insertText works regardless of origin.
• browser_fill_form clears existing value before typing. Use browser_type if you want to append to existing text.
• Check browser_console_messages after interactions to catch JavaScript errors or failed API calls.
• Take a new snapshot after navigation or DOM changes. Refs from old snapshots become stale.
• Use browser_snapshot { "interactive": true } on complex pages to filter to only actionable elements (buttons, links, inputs).
• Use browser_snapshot { "compact": true } to reduce noise from inline text nodes.
• Scroll before screenshotting if the target content is below the fold. Use browser_scroll { "direction": "down" } or browser_scroll { "selector": "#element" } to scroll it into view.

Common Patterns

Login Flow

# 1. Navigate to login page
browser_navigate { "url": "https://app.example.com/login" }

# 2. Snapshot to find form fields
browser_snapshot

# 3. Fill email and password (using refs from snapshot)
browser_fill_form { "ref": "@e3", "value": "[e-mail]" }
browser_fill_form { "ref": "@e4", "value": "mypassword123" }

# 4. Click the login button
browser_click { "ref": "@e5" }

# 5. Wait for navigation to complete
browser_wait_for { "url": "**/dashboard**" }

# 6. Verify with screenshot
browser_screenshot

Form Filling

# 1. Navigate and snapshot
browser_navigate { "url": "https://app.example.com/settings" }
browser_snapshot { "interactive": true }

# 2. Fill each field using refs
browser_fill_form { "ref": "@e2", "value": "John Doe" }
browser_fill_form { "ref": "@e3", "value": "[e-mail]" }
browser_select_option { "ref": "@e4", "values": ["UTC-5"] }

# 3. Submit the form
browser_click { "ref": "@e6" }

# 4. Wait for confirmation
browser_wait_for { "text": "Settings saved" }

Data Extraction

# 1. Navigate to the data page
browser_navigate { "url": "https://example.com/products" }

# 2. Use evaluate to extract structured data
browser_evaluate { "expression": "JSON.stringify([...document.querySelectorAll('.product')].map(el => ({ name: el.querySelector('h2').textContent, price: el.querySelector('.price').textContent })))" }

Visual Testing

# 1. Navigate and set viewport
browser_navigate { "url": "https://example.com" }
browser_resize { "width": 1440, "height": 900 }
browser_screenshot { "fullPage": true }

# 2. Test mobile viewport
browser_resize { "width": 375, "height": 812, "deviceScaleFactor": 3 }
browser_screenshot { "fullPage": true }

Debugging API Calls

# 1. Navigate and perform actions
browser_navigate { "url": "https://app.example.com" }
browser_click { "ref": "@e5" }

# 2. Check network requests for API calls
browser_network_requests { "filter": "*api*", "includeStatic": false }

# 3. Check for console errors
browser_console_messages { "level": "error" }

Waiting for Dynamic Content

# Wait for text to appear after an action
browser_click { "ref": "@e3" }
browser_wait_for { "text": "Results loaded" }

# Wait for a loading spinner to disappear
browser_wait_for { "selector": ".spinner", "state": "hidden" }

# Wait for a custom JS condition
browser_wait_for { "fn": "window.__dataLoaded === true" }

# Wait for URL to change after form submission
browser_wait_for { "url": "**/confirmation**" }

Finding Source Code for UI Elements

# 1. Snapshot the page to identify elements
browser_snapshot

# 2. Inspect a specific element's source
browser_inspect_source { "ref": "@e5" }
# → Component: Hero
# → Source: /src/routes/index.tsx:64:11

# 3. Or use a CSS selector directly
browser_inspect_source { "selector": "nav" }
# → Component: Navbar
# → Source: /src/components/Navbar.tsx:12:5

Working with Multiple Tabs

# List all tabs
browser_tabs

# Open a link in a new tab
browser_click { "ref": "@e10", "newTab": true }

# List tabs again to see the new one
browser_tabs

# Close a specific tab by its target ID
browser_close { "force": true, "targetId": "TARGET_ID_HERE" }