# browser-mcp-cdp

A Model Context Protocol (MCP) server that drives your **real local Chrome** via the Chrome DevTools Protocol. Designed for agents that need access to your logged-in sessions (Gmail, internal dashboards, banking, etc.) without uploading credentials to a cloud service.

## Why another browser MCP?

Different tradeoffs from the alternatives:

| | browser-mcp-cdp | Playwright MCP | Chrome DevTools MCP |
|---|---|---|---|
| Browser | Your local Chrome | Managed Playwright Chromium | Your local Chrome |
| Login state | ✅ Snapshot of your profile (cookies, logins, IndexedDB) | ❌ Fresh profile | ✅ Real profile (modifies it) |
| Profile isolation | ✅ Copy-on-start, real profile untouched | N/A | ⚠️ Direct attach |
| Multi-client | ✅ Shared broker across MCP clients | ❌ Per-client | ❌ Per-client |
| Dependencies | Node only (no Playwright download) | Playwright + browsers | Node only |

The **profile snapshot** pattern means you can ask an agent to "check my email" — it sees your Gmail logged in — but the agent's actions never modify your real Chrome profile. Each session gets a disposable copy.

## Install

```bash
claude mcp add browser -- npx -y browser-mcp-cdp
```

Or add manually to `~/.claude.json` / Claude Desktop config:

```json
{
  "mcpServers": {
    "browser": {
      "command": "npx",
      "args": ["-y", "browser-mcp-cdp"]
    }
  }
}
```

Requires **Node 18+** and Google Chrome installed.

## Tools

| Tool | Purpose |
|---|---|
| `browser_navigate` | Load a URL, return final URL + title |
| `browser_evaluate` | Run JS in the page, return result |
| `browser_screenshot` | PNG of current page |
| `browser_click` | Click via CSS selector |
| `browser_click_at` | Click at pixel coordinates (pierces iframes / overlays) |
| `browser_fill` | Set value on input/textarea/select |
| `browser_wait_for` | Wait for a selector to appear |
| `browser_get_url` | Current URL |
| `browser_get_text` | Visible text of page or element |
| `browser_get_html` | `outerHTML` of page or element |
| `browser_scroll` | Scroll by `(x, y)` pixels |
| `browser_tabs` | List open tabs |
| `browser_new_tab` | Open a new tab |

## Environment variables

| Var | Default |
|---|---|
| `BROWSER_MCP_CHROME_PATH` | Auto-detected per OS |
| `BROWSER_MCP_CHROME_PROFILE_ROOT` | OS-standard Chrome profile dir |
| `BROWSER_MCP_BROKER_SOCKET_PATH` | `~/.browser-mcp/broker.sock` |
| `BROWSER_MCP_BROKER_LOCK_PATH` | `~/.browser-mcp/broker.lock` |
| `BROWSER_MCP_BROKER_IDLE_TIMEOUT_MS` | `600000` (10 min) |

## How it works

1. First MCP client connection spawns a **broker** subprocess (detached).
2. Broker copies your Chrome profile's cookies/local storage/login data to a temp dir.
3. Broker launches Chrome with `--user-data-dir=<temp>` and `--remote-debugging-port`.
4. Subsequent MCP clients connect to the same broker over a Unix socket → they share one Chrome.
5. Broker exits after `BROKER_IDLE_TIMEOUT_MS` of no activity; temp profile is deleted.

## Security

This server lets an LLM execute arbitrary JavaScript in a browser that is logged into your accounts. Only use with models and prompts you trust. Profile snapshotting limits damage (no persistent mutations to your real Chrome), but the agent can read cookies, session tokens, and any data visible to logged-in you for the duration of the session.

## License

MIT