Chrome Debugger MCP

chrome-debugger-mcp

English | 中文

English

An MCP server for breakpoint-driven Chrome debugging.

chrome-debugger-mcp exposes Chrome DevTools Protocol primitives as MCP tools so an AI agent can attach to a real Chrome tab, pause execution, inspect scope values, evaluate expressions inside the current call frame, and step through code with runtime facts instead of guessing from static source.

It is not a general browser automation server. The focus is runtime debugging.

Core Capabilities

• Attach to a real Chrome tab over CDP after explicit user confirmation
• Pause at breakpoints or debugger; statements and wait for the exact pause you expect
• Read local, closure, and module scope values from the paused frame
• Evaluate JavaScript in the current call frame and step execution forward
• Resume cleanly so the agent can continue with actual runtime values

Demo

*Demo: the agent launches Chrome, waits for a breakpoint, inspects real scope variables, and resumes with runtime facts instead of guessing.*

MCP Client Configuration

Use the published package

{

"mcpServers": { "chrome-debugger": { "command": "npx", "args": ["-y", "chrome-debugger-mcp"] } } }

Installation

From npm

npx -y chrome-debugger-mcp

Or install it globally:

npm install -g chrome-debugger-mcp

From source

pnpm install

pnpm build node dist/index.js

Other Highlights

• Launch a dedicated Chrome instance with remote debugging enabled
• Set and remove DevTools breakpoints without editing source code
• Reload the page through CDP so breakpoints reliably bind after navigation
• Poll debugger state when the MCP client has short request timeouts
• Emit _ui payloads and logging messages that clients can surface to the user

Why It Helps

Many browser-focused MCP tools are strong at DOM interaction and network inspection, but weak at runtime debugging. This server gives an MCP client the missing loop you would normally use in Chrome DevTools: attach to the right tab, pause at the right time, inspect real values, step if needed, and resume cleanly.

It also adds guardrails that prevent common agent mistakes:

• guessing which tab to attach to
• concluding behavior without inspecting runtime values
• ending the turn between reloadPage() and waitForSpecificPause()

Requirements

• Google Chrome installed locally
• An MCP client that supports stdio servers and tool calling
• Access to the application you want to debug
• Local source access if you plan to insert temporary debugger; statements

Tooling Model

The server runs over stdio and exposes MCP tools. The most important tools are:

• startDebuggingSession: returns the recommended debugging workflow and critical rules for agent behavior
• launchChrome: launches a dedicated Chrome instance with remote debugging enabled
• listTargets: lists available Chrome tabs and requires the user to pick one
• connect: attaches to the confirmed tab
• setBreakpoint: creates a CDP breakpoint without modifying source files
• removeBreakpoint: removes a breakpoint created by setBreakpoint
• reloadPage: reloads the current page through CDP
• waitForSpecificPause: waits for the next pause and checks whether it matches a target file and line
• waitForPause: waits for any pause without location matching
• getScopeVariables: reads local, closure, and module scope values from the paused frame
• evaluate: executes JavaScript in the paused call frame
• stepInto, stepOver, stepOut: standard execution control
• resume: resumes execution after inspection
• getStatus: non-blocking polling for connected or paused state
• forcePause: requests a pause at the next JavaScript statement

Recommended Workflow

For AI clients, the intended flow is:

1. Call startDebuggingSession().
2. Call launchChrome() or use an already-running Chrome instance with a CDP port.
3. Call listTargets() and show the full tab list to the user.
4. Wait for the user to confirm the exact page URL.
5. Call connect({ targetUrl }).
6. Insert a temporary debugger; statement in local source code, or call setBreakpoint().
7. Call reloadPage().
8. Immediately call waitForSpecificPause() in the same turn.
9. Call getScopeVariables() and evaluate() to inspect runtime values.
10. Step if necessary with stepInto(), stepOver(), or stepOut().
11. Call resume().
12. Remove any temporary debugger; statements from source code.

Important Rules For Agent Authors

This server is designed for tool-using agents, not only for humans. If you are integrating it into an MCP client, keep these rules:

• Never skip listTargets().
• Never guess the target URL, even if only one tab is open.
• Always wait for explicit user confirmation before connect().
• After reloadPage(), immediately call waitForSpecificPause() or waitForPause() in the same turn.
• Do not explain behavior from static code when runtime values can be inspected directly.
• Always resume() after inspection.
• If you added temporary debugger; statements to source code, remove them before finishing.

How waitForSpecificPause Matches

waitForSpecificPause is the preferred waiting primitive because it is more reliable than waiting for an arbitrary pause.

It matches a pause using two strategies:

1. URL fragment plus line tolerance
2. URL fragment plus debugger-statement pause reason

The second path matters when source maps, transpilation, or bundling shift compiled line numbers away from editor line numbers.

Example Tool Sequence

An agent debugging a local Vite app might do something like this:

1. launchChrome({ dryRun: true })
2. launchChrome()
3. listTargets()
4. Wait for the user to confirm http://127.0.0.1:5173
5. connect({ targetUrl: "127.0.0.1:5173" })
6. Insert debugger; in App.jsx
7. reloadPage()
8. waitForSpecificPause({ urlFragment: "App.jsx", line: 62, actionHint: "click the Refetch payloads button" })
9. getScopeVariables()
10. evaluate({ expression: "payload.modules" })
11. resume()

Chrome Launch Behavior

launchChrome() uses a dedicated profile so it does not interfere with the user's normal browser session.

Defaults:

• remote debugging port: 9222
• profile directory: ~/.chrome-debug-profile

Expected Chrome binary locations:

• macOS: /Applications/Google Chrome.app/Contents/MacOS/Google Chrome
• Linux: google-chrome
• Windows: C:\Program Files\Google\Chrome\Application\chrome.exe

If automatic launch fails, the tool returns a command the user can run manually.

Local Playground

This repository includes a disposable test app under test/ so you can exercise the debugger server against a realistic browser workflow.

Start the mock service

cd test/service

node src/server.js

The service listens on http://127.0.0.1:3030.

Start the web app

cd test/web

pnpm install pnpm dev

The web app runs on http://127.0.0.1:5173.

Useful places to pause:

• test/web/src/App.jsx inside loadWorkbench
• test/web/src/App.jsx inside loadModuleDetail
• test/web/src/App.jsx around the unfinished detail sections

Runtime payload areas worth inspecting:

• summaryCards
• modules
• apiContracts
• nextActions
• responseShape

Troubleshooting

No targets found

Make sure Chrome is running with --remote-debugging-port=9222 and the target page is open.

More than one tab matches targetUrl

Pass a more specific substring so the match becomes unique.

waitForPause or waitForSpecificPause times out

This can happen when:

• the page action was never triggered
• the wrong breakpoint was set
• the MCP client itself has a shorter request timeout than the tool call

If your client times out quickly, use getStatus() to poll or increase the client timeout.

The paused line number does not match the editor line

Bundlers and transpilers can shift compiled line numbers. Use waitForSpecificPause() and rely on URL fragment matching plus debugger-statement semantics.

Chrome does not launch automatically

The machine may use a non-default Chrome install path. Run the returned launch command manually or adjust the implementation to match your environment.

Development

pnpm install

pnpm build node dist/index.js

The implementation lives in:

• src/index.ts: MCP tool definitions and user-facing workflow hints
• src/chrome-manager.ts: Chrome DevTools Protocol integration and debugger state management

License

MIT

中文

一个面向 Chrome 断点调试的 MCP Server。

chrome-debugger-mcp 把 Chrome DevTools Protocol 的核心调试能力暴露为 MCP 工具，让 AI agent 可以连接真实的 Chrome 标签页，在运行时暂停执行、读取作用域变量、在当前调用帧中执行表达式、单步跟踪代码，并基于真实值继续任务，而不是只靠静态源码猜测行为。

它不是通用浏览器自动化工具。它的重点是运行时调试。

核心能力

• 在用户明确确认后，通过 CDP 连接真实的 Chrome 标签页
• 在断点或 debugger; 命中时暂停，并等待指定文件和行附近的 pause
• 读取当前暂停帧中的 local、closure、module 作用域变量
• 在当前调用帧里执行 JavaScript，并继续单步跟踪
• 检查完成后恢复执行，让 agent 基于真实运行时值继续工作

功能演示

*演示流程：agent 拉起 Chrome，等待断点命中，读取真实作用域变量，再基于运行时事实继续执行，而不是靠猜测推进。*

MCP 客户端配置

使用已发布包

{

"mcpServers": { "chrome-debugger": { "command": "npx", "args": ["-y", "chrome-debugger-mcp"] } } }

安装方式

从 npm 使用

npx -y chrome-debugger-mcp

也可以全局安装：

npm install -g chrome-debugger-mcp

从源码运行

pnpm install

pnpm build node dist/index.js

其他特点

• 启动带远程调试端口的独立 Chrome 实例
• 无需修改源码即可设置和移除断点
• 通过 CDP 重载页面，确保跳转后断点可靠绑定
• 当 MCP 客户端请求超时较短时，可轮询调试器状态
• 输出 _ui 结果和 logging 消息，方便客户端展示给用户

为什么适合这个场景

很多浏览器方向的 MCP 工具更擅长 DOM 操作和网络请求观察，但不擅长回答运行时调试问题。这个服务补上的是 Chrome DevTools 里最关键的那条链路：连接正确标签页、在正确时机暂停、读取真实值、必要时单步跟踪、最后恢复执行。

它也内置了几条 guardrails，避免 agent 出现这些常见错误：

• 猜测应该连接哪个标签页
• 没看运行时值就直接下结论
• 在 reloadPage() 和 waitForSpecificPause() 之间错误地结束当前轮次

运行要求

• 本机安装了 Google Chrome
• 使用支持 stdio MCP server 和工具调用的 MCP 客户端
• 可以访问你要调试的应用
• 如果要插入临时 debugger;，需要能访问本地源码

工具模型

这个服务通过 stdio 运行，并暴露一组 MCP tools。最核心的工具有：

• startDebuggingSession：返回推荐调试流程和 agent 行为约束
• launchChrome：启动带远程调试能力的独立 Chrome 实例
• listTargets：列出可调试标签页，并要求用户做选择
• connect：连接到已确认的目标标签页
• setBreakpoint：在不改源码的情况下通过 CDP 设置断点
• removeBreakpoint：移除通过 setBreakpoint 创建的断点
• reloadPage：通过 CDP 重载当前页面
• waitForSpecificPause：等待下一次暂停，并判断是否命中目标文件和行
• waitForPause：不做位置匹配，等待任意暂停
• getScopeVariables：读取当前暂停帧中的局部、闭包、模块作用域变量
• evaluate：在暂停调用帧中执行 JavaScript
• stepInto、stepOver、stepOut：标准单步控制
• resume：检查完毕后恢复执行
• getStatus：非阻塞方式查询是否已连接、是否已暂停
• forcePause：请求在下一条 JavaScript 语句处暂停

推荐工作流

对于 AI 客户端，建议流程是：

1. 调用 startDebuggingSession()。
2. 调用 launchChrome()，或直接复用已经开启 CDP 端口的 Chrome。
3. 调用 listTargets()，并把完整标签页列表展示给用户。
4. 等待用户明确确认要调试的页面 URL。
5. 调用 connect({ targetUrl })。
6. 在本地源码插入临时 debugger;，或者调用 setBreakpoint()。
7. 调用 reloadPage()。
8. 在同一轮里立刻调用 waitForSpecificPause()。
9. 调用 getScopeVariables() 和 evaluate() 检查运行时值。
10. 必要时使用 stepInto()、stepOver()、stepOut() 继续跟踪。
11. 调用 resume()。
12. 删除源码里临时加入的 debugger;。

给 Agent 作者的重要规则

这个服务首先是为会调用工具的 agent 设计的，而不仅仅是给人手动点工具用。如果你要把它接入自己的 MCP 客户端，建议遵守这些规则：

• 不要跳过 listTargets()。
• 即使只看到一个标签页，也不要猜测目标 URL。
• 一定要等用户明确确认后再调用 connect()。
• 调用 reloadPage() 后，必须在同一轮里立刻调用 waitForSpecificPause() 或 waitForPause()。
• 能读取运行时值时，不要只根据静态代码解释行为。
• 检查完之后一定要 resume()。
• 如果向源码里插入了临时 debugger;，结束前要清理掉。

waitForSpecificPause 如何匹配

waitForSpecificPause 是首选的等待工具，因为它比“等待任意暂停”更可靠。

它有两层匹配策略：

1. URL 片段加行号容差
2. URL 片段加 debugger-statement 暂停原因

第二层匹配对经过 source map、转译、打包后的代码尤其重要，因为编译后的行号可能和编辑器行号不完全一致。

调用序列示例