io.github.xianguoGou/h5-devtools-mcp
Debug H5 pages inside Android app WebViews: console logs, exceptions, run JS.
Installation
Ajoutez ceci à la configuration MCP de votre client (Claude Desktop : claude_desktop_config.json ; Cursor : .cursor/mcp.json ; VS Code et les autres utilisent la même forme « mcpServers »).
h5-devtools-mcpclaude mcp add h5-devtools-mcp -- npx -y h5-devtools-mcp@0.1.2
{
"mcpServers": {
"h5-devtools-mcp": {
"command": "npx",
"args": [
"-y",
"h5-devtools-mcp@0.1.2"
]
}
}
}
Documentation
Tirée du README du projet sur GitHub, actualisée chaque semaine. Les liens ouvrent GitHub.
Sommaire
h5-devtools-mcp
English | 简体中文
An MCP server that lets coding agents (Claude Code, Codex, Cursor…) debug H5 pages running inside Android WebViews on real devices — read console output and uncaught exceptions, and run JavaScript in the page, without opening chrome://inspect.
Status: v0.1, early. Android only.
How it works
Coding agent ──MCP (stdio)──▶ h5-devtools-mcp ──adb forward──▶ webview_devtools_remote_<pid> ──CDP──▶ WebView page
Every debuggable WebView exposes a Chrome DevTools Protocol endpoint on an abstract unix socket. The server discovers those sockets over adb (the same way chrome://inspect does), forwards them to local ports, and connects with CDP.
Requirements
- Node.js 18+
adbon yourPATH(or setADB_PATH)- An Android device with USB debugging enabled
- The app must enable WebView debugging (debug builds usually do):
WebView.setWebContentsDebuggingEnabled(true);
Setup
# Claude Code
claude mcp add h5-devtools -- npx -y h5-devtools-mcp
Other MCP clients (Cursor, Codex…) take the same command: npx -y h5-devtools-mcp.
From source
git clone https://github.com/xianguoGou/h5-devtools-mcp.git
cd h5-devtools-mcp
npm install && npm run build
claude mcp add h5-devtools -- node /absolute/path/to/h5-devtools-mcp/dist/index.js
Tools
| Tool | What it does |
|---|---|
list_targets | List debuggable pages: app package, title, URL, visibility |
attach | Connect to a page and start buffering console output |
get_console | Read buffered entries, filtered by level / keyword / exclude_keywords, incrementally via since_seq |
evaluate_js | Run a JavaScript expression in the page and return the result (promises are awaited) |
detach | Disconnect from a page |
Output is designed to be token-cheap: compact one-line entries, long messages truncated, and since_seq so an agent only reads what's new instead of re-reading the whole buffer on every turn.
Example
> list_targets
- id: R5CT1234/webview_devtools_remote_8812/3A1F…
app: com.example.app (pid 8812) on SM-S9180
title: "Order"
url: https://m.example.com/order
> get_console { "levels": ["error"] }
#41 10:32:01.123 [error exception] TypeError: Cannot read properties of undefined (reading 'price')
at https://m.example.com/assets/index-8f2a.js:1:48213
next since_seq: 57
Troubleshooting
- "no debuggable WebView found" — the app hasn't enabled WebView debugging, or no WebView is alive yet. Open the H5 page first.
- "debugger in use" — only one debugger can attach to a page. Close the
chrome://inspectDevTools window. - "unauthorized" — accept the USB debugging prompt on the device.
Known limitations
- Debugger-in-use detection is best-effort. It relies on
/json/listomittingwebSocketDebuggerUrl. Recent Chromium accepts several CDP clients per page, so attach may succeed alongside an open DevTools window; older WebViews may still refuse. list_targetscan take ~3s per unresponsive app. Sockets are queried one by one with a 3s timeout, and Android freezes background apps, so their WebView sockets stop answering.attachonly re-checks the target's own socket and stays fast.evaluate_jsruns with the page's full privileges — it can read cookies, storage and tokens, and change page state. Keep it behind your agent's tool-permission prompt.- The e2e script expects port 9333 to be free. If a previous run left a headless Chromium behind, it attaches to that one; kill it first.
Development
The e2e test uses a fake adb and headless Chromium in place of a phone:
npm run build
CHROME=/path/to/chrome npm run test:e2e
Roadmap
- [x] Discover devices and WebView pages
- [x] Console messages, uncaught exceptions, browser log entries
- [ ] Network requests and responses
- [ ] JSBridge call logging (configurable bridge shapes)
- [x]
evaluate_js - [ ]
screenshot - [ ] JSBridge mocking
- [ ] iOS (WKWebView via WebKit Inspector Protocol)
License
MIT
Questions fréquentes
Qu'est-ce que io.github.xianguoGou/h5-devtools-mcp ?
Debug H5 pages inside Android app WebViews: console logs, exceptions, run JS.
Comment installer io.github.xianguoGou/h5-devtools-mcp ?
Il s'exécute sur votre machine. Dans Claude Code : claude mcp add h5-devtools-mcp -- npx -y h5-devtools-mcp@0.1.2. Pour les autres clients, collez la configuration JSON de cette page dans leurs réglages MCP.
Faut-il une clé API pour io.github.xianguoGou/h5-devtools-mcp ?
Son entrée au registre n'indique aucune clé ni secret obligatoire. Certaines fonctions peuvent tout de même demander un compte auprès du service auquel il se connecte.
io.github.xianguoGou/h5-devtools-mcp est-il open source ?
Oui. Son code est sur GitHub sous licence MIT.
Qui maintient io.github.xianguoGou/h5-devtools-mcp ?
Il est publié par xianguoGou. AI.tn le référence sans le maintenir ni le recommander.
Issu du registre MCP officiel. Être listé n'est pas une recommandation : vérifiez le code et les permissions d'un serveur avant de lui donner accès à vos données.