Chrome扩展测试MCP服务

Chrome Extension Tester — MCP Server

An MCP (Model Context Protocol) server that lets Claude interactively test any unpacked Chrome extension using Playwright. Load your extension, interact with its popup and options page, inspect storage, monitor network requests, check badges, test messaging, and more — all through natural language.

Table of Contents

• Why
• Architecture
• Features
• Requirements
• Installation
• Setup with Claude Desktop
• Setup with Claude Code (npx)
• Available Tools
• Testing Agent Prompt
• Example: testing an extension popup
• Example: testing against your real browser
• Example Prompts
• Project Structure
• Notes

Why

Testing a Chrome extension during development means manually clicking reload, opening the popup, checking storage in DevTools, watching the service worker console, and copy-pasting errors back to the agent on every iteration. This MCP server gives an AI coding agent direct access to all of that through tool calls, so the agent can iterate on its own. It exists because the manual loop made working with Claude Code on extensions too slow.

Architecture

graph LR
    A[AI Agent<br/>Claude / Cursor] -->|MCP protocol| B[This server<br/>17 tools]
    B -->|Playwright: load_extension| C[Chromium<br/>launched context]
    B -->|Playwright CDP: connect_browser| D[Your real Brave/Chrome<br/>with existing logins]
    C -->|loads| E[Extension under test]
    D -->|attaches to| E
    B -.->|reads / writes| F[(state.js<br/>browser, context, page,<br/>extensionId, connectionMode)]

Two connection modes:

• CDP (default/preferred) — connect_browser attaches to your running Brave or Chrome via the Chrome DevTools Protocol. All your existing logins and tabs are preserved, so you can test against real authenticated sessions. Use this whenever the extension interacts with a real site that requires being logged in. reload_extension hot-reloads code changes in this mode without restarting the browser.
• Launched (fallback) — load_extension spins up a fresh, isolated Chromium with only your extension installed and no real logins. Use this only for self-contained extensions (on-device models, local-only tools) that don't depend on a real account/session. Refuses to run while a connect_browser session is active, since it would tear that session down.

Features

• Connect to your real Brave or Chrome with all existing logins intact — test against authenticated sessions without re-logging in; this is the default/preferred way to test
• Hot-reload extension code in your real browser via reload_extension — no restart, tabs and logins survive
• List installed extensions by name and target any one by name or ID; throws on ambiguous matches instead of silently picking the wrong extension
• Load and reload any unpacked Chrome extension in an isolated Chromium instance (for self-contained extensions with no real-account dependency)
• Interact with popup and options pages (click, type, read content)
• Inspect and manipulate chrome.storage (local / sync / session) — always targets the correct extension's service worker, never a neighbouring extension's
• Read background service worker logs — filtered to your target extension only
• Monitor and inspect network requests
• Check and assert badge text and color
• Send messages to the background script and validate responses
• Enumerate, open, close, and switch browser tabs — works in both launched and real-browser (CDP) mode
• Test context menu registration and handler invocation
• Run assertions that return structured PASS / FAIL results
• Take screenshots at any point during testing
• Create and reuse test accounts on any website using disposable email (via Guerrilla Mail API)
• Inspect, set, delete, and assert browser cookies — including httpOnly cookies invisible to document.cookie

Requirements

• Node.js 18 or higher
• Claude Desktop or Claude Code with MCP support
• A Chrome extension with a manifest.json (Manifest V2 or V3)

Installation

Option A — npx (no install needed)

npx chrome-extension-tester-mcp

Option B — install globally

npm install -g chrome-extension-tester-mcp

Option C — clone and run locally

git clone https://github.com/BHUVAN-RJ/chrome-extension-testing-mcp.git
cd chrome-extension-testing-mcp
npm install
npx playwright install chromium

Setup with Claude Desktop

Add the following to your Claude Desktop MCP config file:

macOS / Linux — ~/.config/claude/claude_desktop_config.json Windows — %APPDATA%\Claude\claude_desktop_config.json

Using npx (recommended)

{
  "mcpServers": {
    "chrome-extension-tester": {
      "command": "npx",
      "args": ["chrome-extension-tester-mcp"]
    }
  }
}

Using a local clone

{
  "mcpServers": {
    "chrome-extension-tester": {
      "command": "node",
      "args": ["/absolute/path/to/chrome-extension-testing-mcp/src/index.js"]
    }
  }
}

Restart Claude Desktop after saving the config.

Setup with Claude Code (npx)

Add to your project's .mcp.json or user-level MCP config:

{
  "mcpServers": {
    "chrome-extension-tester": {
      "command": "npx",
      "args": ["chrome-extension-tester-mcp"]
    }
  }
}

Available Tools

connect_browser

Attach to your real Brave or Chrome browser, preserving all existing logins and open tabs.

Inputs:

• action (string, required): scan | connect | launch | list_extensions
• port (number, default 9222) — CDP debug port, for connect
• browser_name (string, default "Brave") — which browser to launch, for launch
• debug_port (number, default 9222) — debug port to use when launching, for launch
• extension_id (string) — 32-char extension ID or a name substring (e.g. "MyExt") to target for connect / launch. Throws if the name matches more than one installed extension — use the full ID in that case.
• browser_name_for_extensions (string, default "Brave") — which browser's extension directory to scan, for list_extensions

Actions:

Returns: Connection confirmation with the resolved extension ID, or the list of extensions/browsers.

Note: To use connect, your browser must be started with --remote-debugging-port=9222. Use launch to have the server start it for you. Only one Chrome-based browser instance can run per profile — close your existing window before using launch.

load_extension

Fallback for self-contained extensions with no real-account dependency. Launch an isolated Chromium with an unpacked extension and capture its ID. Inputs: extension_path (string, required) — path to the unpacked extension folder. Returns: Text confirming the resolved path and the detected extension ID. Throws: if a connect_browser (CDP) session is active — use reload_extension instead so the real browser session isn't torn down.

reload_extension

Hot-reload the extension's code in the connected real browser (connect_browser mode only), without restarting the browser. Clicks the "Update" button on chrome://extensions and waits 10s for it to settle. Inputs: none. Returns: Text confirming the click and extension ID (unchanged — tabs/storage survive). Note: reloads every unpacked extension in the profile and update-checks every Web Store one — not scoped to a single extension.

interact_with_popup

Open the popup and click, type, or read its DOM. Inputs: action (string, required: open | click | type | get_text | get_html); selector (string); value (string, for type). Returns: Text describing the action result, or the requested text/HTML.

open_options_page

Open the options page (or any extension page) and interact with it. Inputs: page (string, default options.html); action (string: open | click | type | get_text | get_html); selector (string); value (string). Returns: Text describing the action result, or the requested text/HTML.

inspect_dom

Query a selector or evaluate JS in a page, optionally navigating first. Inputs: url (string); selector (string); script (string, overrides selector). Returns: Text with matched elements' outerHTML, or the JSON-serialized script result.

In CDP mode, if the active page is a restricted URL (devtools://, chrome://, chrome-extension://), a fresh tab is opened automatically rather than failing.

get_service_worker_logs

Read buffered console logs from the target extension's background service worker. Inputs: clear_after (boolean, default false). Returns: Text listing captured log entries, or a "none captured yet" message.

Logs are filtered to the extension set via extension_id in connect_browser. In a real browser with many extensions installed, only logs from your target extension are captured — not those from other extensions like Grammarly or password managers.

take_screenshot

Save a screenshot of the current page or popup. Inputs: output_path (string, default ./screenshot.png); full_page (boolean, default false). Returns: Text with the saved file path.

run_assertion

Assert an element exists/has text, or that a JS expression is truthy. Inputs: description (string, required); selector (string); expected_text (string); script (string, overrides selector). Returns: Text beginning with PASS or FAIL, followed by detail.

extension_storage

Read from or write to chrome.storage (local / sync / session). Inputs: action (string, required: get | set | remove | clear); area (string, default local); keys (string[]); data (object, for set). Returns: Text with storage contents, or a confirmation of the write/removal.

Operates via the target extension's service worker. If no service worker is found for the targeted extension ID, throws a descriptive error listing all active workers so you can retarget with connect_browser.

monitor_network

Capture and inspect network requests during navigation. Inputs: action (string, required: navigate_and_capture | get_captured | clear); url (string); filter_pattern (string); include_types (string[]). Returns: Text listing captured requests as [method] [type] status url.

check_badge

Read or assert the action badge text and background color. Inputs: action (string, required: get | assert_text | assert_color); tab_id (number); expected_text (string); expected_color (number[] RGBA). Returns: Text with the badge value, or a PASS / FAIL assertion result.

send_message_to_background

Send chrome.runtime.sendMessage from the popup and return the response. Inputs: message (object, required); timeout_ms (number, default 5000). Returns: Text with the sent message and JSON response, or a failure message.

test_context_menu

Check the contextMenus API, simulate a right-click, or trigger an item. Inputs: action (string, required: check_api | right_click | trigger_item); url (string); selector (string); menu_item_id (string); page_url (string). Returns: Text with API availability, dispatch confirmation, or trigger result.

simulate_tab_events

Open, close, switch, list, or close all browser tabs. Inputs: action (string, required: open | close | switch | list | close_all); url (string); tab_index (number). Returns: Text describing the affected tab(s), or the list of open tabs.

Works in both launched Chromium and CDP (real browser) mode. In CDP mode, lists and controls all tabs in your real browser, including ones opened by extension automation.

test_account_login

Create or reuse a test account on a site using a disposable email. Inputs: action (string, required: auto | create | login); account_key (string, required); signup_url / login_url (string); selector overrides (email_selector, password_selector, submit_selector, pre_click_selector); multi-step fields (step2_url, step2_password_selector, step2_submit_selector). Returns: Text reporting account creation/login status plus a screenshot path.

manage_cookies

Inspect, edit, delete, or assert browser cookies (operates on the Playwright browser context, so it works regardless of the extension's permissions and can read httpOnly cookies). Works in both connection modes. Inputs: action (string, required: get | set | delete | clear | assert); url (string — scope for get/assert, or the cookie URL for set); name (string — filter for get, target for delete, required for assert); cookie (object, for set: name, value, url or domain+path, plus optional expires, httpOnly, secure, sameSite); domain / path (string filters for delete, and disambiguators for assert); expected_value (string, for assert); include_values (boolean, for get and assert in CDP mode); overwrite (boolean, for set in CDP mode). Returns: JSON-serialized cookies for get; a confirmation for set/clear; a count plus the name @ domain/path of every removed cookie for delete (or an explicit "nothing deleted" when the filter matched none); a PASS / FAIL line for assert.

Domain matching. A leading dot is ignored on both sides, so domain: "example.com" matches a cookie stored as .example.com — the form a server's Set-Cookie; Domain= produces. delete then requires exact domain identity, so it never reaches subdomains; assert additionally matches a parent domain-cookie that would be sent to the domain you name (.example.com satisfies domain: "app.example.com", a host-only example.com cookie does not). path is matched exactly, in both delete and assert.

Ambiguous asserts. One cookie name can exist at several domain/path slots (sid at / and at /admin). If more than one candidate matches and they hold different values, assert refuses rather than picking one by cookie-store order, and lists the slots (names and paths only, never values) so you can narrow it with domain and/or path. Candidates that agree on a value assert normally.

Guardrails in connect_browser (CDP) mode. There the cookies are your real profile's credentials, not throwaway test data, so the tool is deliberately restricted:

Testing Agent Prompt

The server includes a built-in MCP prompt called extension-tester-agent — a fully automated testing agent that validates all implemented changes and returns a structured report.

Arguments

What it does

1. Chooses a browser mode — defaults to connect_browser (real logged-in browser); only uses load_extension if the extension is self-contained with no real-account dependency
2. Understands the extension and derives a set of tests from the changes list
3. Writes a test plan — every change maps to at least one test and the right MCP tool
4. Executes every test — never skips, takes screenshots on failure
5. Reports a structured PASS / FAIL table with details on any failures

How to invoke

After implementing changes, tell Claude:

Use the extension-tester-agent prompt with:
- extension_path: /path/to/my-extension
- extension_description: "A tab manager that saves sessions to chrome.storage.local and restores them via a popup"
- changes: "Added save button; save button writes open tabs to storage.local; badge shows count of saved tabs"

Claude will write the test plan, execute every test, and return a full report.

Example: testing an extension popup

A typical loop the agent can run on its own:

1. Load the extension

{ "tool": "load_extension", "arguments": { "extension_path": "/tmp/my-extension" } }

Extension loaded.
Path: /tmp/my-extension
Extension ID: ddnjmkpjnchafihagpljebmkdpejhaoj

2. Open the popup and read its HTML

{ "tool": "interact_with_popup", "arguments": { "action": "open" } }

<body>
  <h1>Tab Saver</h1>
  <button id="save">Save open tabs</button>
  <span id="count">0 saved</span>
</body>

3. Read local storage

{ "tool": "extension_storage", "arguments": { "action": "get", "area": "local" } }

storage.local contents:
{
  "savedTabs": [],
  "enabled": true
}

Example: testing against your real browser

Use this flow when your extension needs to interact with pages where you're already logged in (LinkedIn, Gmail, internal tools, etc.).

1. Start Brave with remote debugging (one-time setup)

/Applications/Brave\ Browser.app/Contents/MacOS/Brave\ Browser --remote-debugging-port=9222

Or just use launch to have the server handle it:

2. Scan and connect

{ "tool": "connect_browser", "arguments": { "action": "scan" } }

Running browsers with remote debugging:
  Port 9222: Chrome/149.0.7827.54

Use action:"connect" with port:9222 to attach.

3. List installed extensions and pick the right one

{ "tool": "connect_browser", "arguments": { "action": "list_extensions" } }

Installed extensions (5):
  eimadpbcbfnmbkopoojfekhnkhdbieeh  Dark Reader
  ghmbeldphafepmbegfdlkpapadhbakde  Proton Pass: Free Password Manager
  ikiaabefahnnfjninfhfeolbgkfnldpa  MyJobBot - Autofill
  ...

Pass the ID or name substring as extension_id when connecting.

4. Connect targeting your extension by name

{ "tool": "connect_browser", "arguments": { "action": "connect", "port": 9222, "extension_id": "MyJobBot" } }

Connected to Chrome/149.0.7827.54 on port 9222.
Extension ID: ikiaabefahnnfjninfhfeolbgkfnldpa
Your existing tabs and logged-in sessions are preserved.

5. List tabs, switch to one opened by the extension, inspect the DOM

{ "tool": "simulate_tab_events", "arguments": { "action": "list" } }
{ "tool": "simulate_tab_events", "arguments": { "action": "switch", "tab_index": 4 } }
{ "tool": "inspect_dom", "arguments": { "script": "document.querySelector('#apply-button')?.textContent" } }

6. Made a code change? Hot-reload without losing the session

{ "tool": "reload_extension", "arguments": {} }

Clicked 'Update' on chrome://extensions and waited 10s for it to settle.
Extension ID: ikiaabefahnnfjninfhfeolbgkfnldpa (unchanged — tabs and chrome.storage data should survive).

Example Prompts

Load my extension from /Users/me/my-extension and open the popup

Click the button with selector #save and take a screenshot

Navigate to https://example.com and check if my content script injected a .banner element

Read all keys from chrome.storage.local

Set { "enabled": true } in chrome.storage.local and verify it was saved

Navigate to https://example.com, capture all network requests, then show me any that were blocked

Check the badge text — it should say "ON"

Send the message { "type": "GET_STATUS" } to the background and show the response

Open a tab to https://news.ycombinator.com, then another to https://github.com, then list all open tabs

Connect to my real Brave browser and list all installed extensions

Connect to Brave targeting the "MyJobBot" extension and read its chrome.storage.local

List all open tabs in my Brave browser and switch to the LinkedIn one

Right-click on https://example.com and trigger the context menu item with id "my-action"

Create a test account on https://example.com/signup and save it as "my_test_account"

Log in to https://example.com/login using the stored "my_test_account" credentials

Project Structure

chrome-extension-testing-mcp/
├── src/
│   ├── index.js              # MCP server entry point
│   ├── state.js              # Shared browser state and helpers (ensureBrowser, ensurePage, getServiceWorker)
│   ├── prompts/
│   │   ├── index.js          # Registers MCP prompts
│   │   └── extension-tester.js  # extension-tester-agent prompt definition
│   └── tools/
│       ├── index.js          # Aggregates all tool definitions and handlers
│       ├── connect-browser.js  # CDP connection to real Brave/Chrome; list_extensions; name-based targeting
│       ├── reload-extension.js # Hot-reload via chrome://extensions "Update" button (CDP mode only)
│       ├── load-extension.js
│       ├── popup.js
│       ├── dom.js
│       ├── logs.js
│       ├── screenshot.js
│       ├── assertion.js
│       ├── storage.js
│       ├── network.js
│       ├── options-page.js
│       ├── context-menu.js
│       ├── badge.js
│       ├── messaging.js
│       ├── tabs.js
│       └── account-login.js
├── package.json
└── README.md

Notes

• The browser launches in headed mode (visible window) so you can watch tests run in real time
• Screenshots default to ./screenshot.png unless a custom path is provided
• Service worker logs are buffered from the moment the browser connects; in CDP mode they are filtered to the targeted extension only
• Call load_extension again at any time to get a fresh isolated browser instance
• connect_browser is the default/preferred way to test — use load_extension only for extensions with no real-account dependency
• reload_extension reloads every unpacked extension in the profile (and update-checks Web Store ones) as a side effect — not scoped to a single extension
• Native Chrome context menus cannot be automated by Playwright — use test_context_menu with trigger_item to invoke handlers directly
• Badge and storage tools communicate via the service worker, so the extension must have a background service worker (MV3)
• test_account_login uses the Guerrilla Mail API to generate disposable emails — no browser navigation required, no bot-blocking. Credentials are stored in test-accounts.json at the project root (add this to .gitignore)
• Use action: "auto" for test_account_login to automatically reuse stored credentials or create a new account if none exist
• In CDP mode (connect_browser), inspect_dom automatically avoids navigating restricted pages (devtools://, chrome://, chrome-extension://) by opening a new tab when needed
• Only one Chrome-based browser instance can run per profile — if connect_browser launch fails, close the existing browser window first
• extension_id in connect_browser accepts either a full 32-character ID or a name substring. If the substring matches multiple extensions, an error is thrown listing all matches so you can switch to the full ID

License

MIT