Provides full browser control including navigation, element interaction, form filling, screenshot capture, DevTools inspection (console, network, elements, storage), tab management, and multi-session isolation via Chrome's remote debugging protocol.
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., "@chrome-agent-mcp go to example.com and take a screenshot"
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.
Full Chrome control for AI agents via the Model Context Protocol
Give Claude (or any MCP client) full control over your browser: navigate pages, click elements, fill forms, inspect DevTools — with parallel tab execution, session isolation per agent, and a real-time dashboard.
Each Claude session gets its own Chrome tab group — so multiple agents can run in parallel without stepping on each other. A companion Chrome extension shows a live dashboard: active tabs, event log, and screenshots in real time.
Claude session A → tab group "Pinguim" (blue)
Claude session B → tab group "Fenix" (red)
Claude session C → tab group "Girafa" (yellow)Every tool call is wrapped with resilient execution:
20 s primary timeout
2× parallel retries on timeout (10 s each)
Automatic fallback to a new Chrome group if all retries fail
Related MCP server: OpenChrome
Node.js ≥ 18
Google Chrome with remote debugging enabled
google-chrome --remote-debugging-port=9222To make it permanent, add the flag to your Chrome launcher or .desktop file.
# Run directly (no install)
npx chrome-agent-mcp
# Or install globally
npm install -g chrome-agent-mcpClaude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"chrome-agent-mcp": {
"command": "npx",
"args": ["chrome-agent-mcp"],
"env": { "CHROME_DEBUG_PORT": "9222" }
}
}
}Claude Code (.claude/settings.json):
{
"mcpServers": {
"chrome-agent-mcp": {
"command": "node",
"args": ["/path/to/chrome-agent-mcp/dist/index.js"],
"env": { "CHROME_DEBUG_PORT": "9222" }
}
}
}Open chrome://extensions/
Enable Developer mode
Click Load unpacked → select the extension/ folder
The extension adds a popup dashboard showing which tabs are active, what the agent is doing, and captured screenshots — all updating in real time.
Tool | Description |
| Navigate to a URL |
| Go back in browser history |
| Go forward in browser history |
| Reload the current tab |
Tool | Description |
| Click an element by accessibility ref |
| Type text into a focused element |
| Hover over an element |
| Press a keyboard key or shortcut |
| Scroll the page or an element |
| Select a |
| Fill multiple form fields at once |
| Wait for an element or condition |
Tool | Description |
| Get the accessibility tree of the page |
| Capture a full-page screenshot |
| Execute JavaScript and return the result |
Tool | Description |
| Open, close, switch, list, and mark tabs done |
Tool | Description |
| Read and clear browser console logs |
| Inspect network requests and responses |
| Inspect and query DOM elements |
| Read cookies, localStorage, sessionStorage |
Tool | Description |
| List all open Chrome windows |
| Focus a specific window |
| List installed Chrome extensions |
| Sync session state with the extension dashboard |
src/
├── index.ts MCP server + request dispatch
├── types.ts Shared type definitions
├── core/ Infrastructure layer
│ ├── connection.ts Chrome CDP connection lifecycle
│ ├── bridge.ts Extension event bridge
│ ├── favicon.ts Tab favicon animation manager
│ ├── resilience.ts Timeout / retry / fallback execution
│ └── groups/ Tab group isolation
│ ├── manager.ts Orchestrator (public API)
│ ├── state.ts File-based state persistence
│ └── chrome-api.ts Chrome Extension API bridge
├── tools/ MCP tool implementations
│ ├── tabs.ts
│ ├── navigation.ts
│ ├── browser.ts
│ ├── media.ts screenshot · snapshot · evaluate
│ ├── session.ts
│ ├── interaction/ click · type · hover · scroll · form · wait
│ └── devtools/ console · network · elements · storage
└── utils/ Pure utility functions
├── accessibility.ts AX tree formatter
├── description.ts Tab action labels
└── identity.ts Animal names + Chrome group colorsEngineering rules: every file ≤ 300 lines · every directory ≤ 7 items · SOLID principles throughout.
Variable | Default | Description |
|
| Chrome remote debugging port |
# Install dependencies
npm install
# Build TypeScript
npm run build
# Watch mode
npm run dev
# Run tests
npm test
# Coverage report (target: 80%)
npm run coverageMIT © Otavio Senne
Click on an element using its ref from browser_snapshot. Supports targeting a specific tab via tabId.
Inputs: ref (required), tabId, agentId, element, +1 more.
Evaluate JavaScript in a page context. IMPORTANT: top-level const/let/var declarations persist across calls — always wrap multi-statement code in an IIFE: (() => { const x = ...; return x; })(). Never use bare 'return' at top level. Max timeout: 10s — avoid setTimeout > 10000ms. Supports targeting a specific tab via tabId.
Inputs: expression (required), tabId, agentId.
Fill multiple form fields at once. Supports targeting a specific tab via tabId.
Inputs: fields (required), tabId, agentId.
Hover the mouse over an element. Supports targeting a specific tab via tabId.
Inputs: ref (required), tabId, agentId, element.
Press a keyboard key in the specified tab.
Inputs: key (required), tabId, agentId.
Reload the page in the specified tab
Inputs: tabId, agentId.
Scroll the page or a specific element. Supports targeting a specific tab via tabId.
Inputs: direction (required), ref, tabId, amount, +1 more.
Select an option in a element. Supports targeting a specific tab via tabId.
Inputs: ref (required), value (required), tabId, agentId, +1 more.
Capture accessibility snapshot of a page. Returns the page structure with element refs for interaction. Use instead of screenshots to find elements to click/type. IMPORTANT: refs go stale after any DOM change or navigation — always re-snapshot before clicking if the page may have changed. Supports targeting a specific tab via tabId.
Inputs: tabId, agentId.
List, create, close, or switch browser tabs. Returns tabId for each tab — pass tabId to other tools to target a specific tab without switching.
Inputs: action (required), url, index, tabId, +1 more.
Take a screenshot of a page. Use browser_snapshot to find interactable elements instead. Supports targeting a specific tab via tabId.
Inputs: tabId, agentId, fullPage.
Type text into an input element. Supports targeting a specific tab via tabId.
Inputs: ref (required), text (required), tabId, submit, +2 more.
Wait for text to appear on the page or wait a fixed duration. Max 10s per call — call multiple times for longer waits. If a previous attempt failed, do NOT increase wait time blindly; try a different action instead. Supports targeting a specific tab via tabId.
Inputs: text, time, tabId, agentId.
List installed Chrome extensions.
Inputs: no named parameters.
Bring a specific Chrome tab into focus (activate it in the browser UI).
Inputs: tabId (required).
Manage Chrome browser windows. List all windows with their tabs, or focus a specific window. Each window groups multiple tabs.
Inputs: action (required).
Inspect the browser console for a tab. Start monitoring to capture future logs, then list or clear them. Equivalent to the DevTools Console panel.
Inputs: action (required), level, limit, tabId, +1 more.
Inspect DOM elements and computed CSS styles for a tab. Equivalent to the DevTools Elements panel.
Inputs: selector (required), tabId, agentId, includeStyles.
Inspect network requests for a tab. Start monitoring to capture future requests, then list or clear them. Equivalent to the DevTools Network panel.
Inputs: action (required), limit, tabId, filter, +2 more.
Read or write browser storage for a tab: localStorage, sessionStorage, or cookies. Equivalent to DevTools Application > Storage panel.
Inputs: action (required), storageType (required), key, tabId, +2 more.
Sync Claude session name/color with Chrome tab group. 'rename': renames BOTH the Claude chat AND Chrome group automatically. 'color': maps a Claude Code color to Chrome group color and returns the /color command. 'detect': reads current Chrome group state and returns matching /rename + /color commands.
Inputs: action (required), claudeColor, description.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 24 tools
Most tools are clearly distinct action primitives (click, type, scroll, navigate, etc.), but there is overlap among browser_tabs, chrome_windows, and chrome_focus around tab/window switching. browser_snapshot and devtools_elements both expose page structure, though one is interaction-oriented and the other is inspection-oriented.
The tool names mostly follow a clear browser_<verb>, devtools_<panel>, or chrome_<noun> pattern, which makes the set predictable. browser_tabs is a noun-based multi-action tool rather than a verb-noun name, and session_sync breaks the prefix convention.
24 tools covers a broad browser-automation surface, but it is on the heavy end and includes very granular actions like navigate_back and navigate_forward that could have been combined. The count is justified by the inclusion of devtools and window management, yet it feels more like a toolkit than a focused MCP server.
The server covers core browser workflows well: tab lifecycle, navigation, page interaction, devtools inspection, storage, and window management. Minor gaps exist such as advanced network interception and drag-and-drop, but these can often be worked around with browser_evaluate.
Give AI agents a real cloud Chrome: open sessions, navigate, read pages, click, type and screenshot.
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
Let AI agents use your real, signed-in Chrome: tabs, clicks, forms, screenshots and video.
Stealth web automation for AI agents. Login, signup, navigate, screenshot.