Helm Browser

Browser control for AI coding assistants via MCP. Your real Chrome, not a headless instance.

npx skills add jfikrat/helm-browser

Highlights

• ~57 tools across navigation, reading, interaction, scripting, state, and multi-target control
• Persistent debugger sessions with per-tab serialization for debugger-mutating commands
• Semantic locators via browser_get_snapshot() and locator support for stable element targeting
• Multi-target support for iframes, popups, and dialogs

Why Helm?

Most browser automation tools launch a separate headless Chromium. Helm takes a different approach: it controls your actual Chrome through a lightweight extension.

Supported Clients

• Claude Code - Anthropic's CLI
• Codex CLI - OpenAI
• Gemini CLI - Google
• Aider - AI pair programming
• Any MCP-compatible client

Architecture

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   AI Client     │────▶│   Helm Daemon   │────▶│ Chrome Extension│
│ (Claude, etc.)  │ WS  │   (Port 9876)   │ WS  │   (Side Panel)  │
└─────────────────┘     └─────────────────┘     └─────────────────┘

• Extension: Chrome extension with side panel UI
• Daemon: WebSocket coordinator on port 9876
• Server: MCP server exposing browser tools
• Client: MCP client for AI assistants

Quick Start

Prerequisites

• Chrome or Chromium

1. Install

git clone https://github.com/jfikrat/helm-browser.git
cd helm-browser
./install.sh

2. Load Chrome Extension

1. Go to chrome://extensions
2. Enable Developer mode
3. Click Load unpacked → select extension/ folder

install.sh installs Bun if needed, runs bun install in daemon/, server/, and client/, and registers the daemon as a user service (systemd on Linux, launchd on macOS). It is safe to re-run.

To remove the service and installed dependencies later:

./uninstall.sh

3. Configure MCP

Add to your MCP config (e.g., ~/.claude/claude.json):

{
  "mcpServers": {
    "helm": {
      "command": "bun",
      "args": ["run", "/path/to/helm-browser/client/index.ts"]
    }
  }
}

Tools

Navigation

Reading

Interaction

Targets & Debugging

Advanced

Examples

Web Scraping

> Navigate to Hacker News, get the top 3 story titles

browser_navigate("https://news.ycombinator.com")
browser_get_element_text(".titleline > a")

Form Automation

> Log into my dashboard

browser_navigate("https://app.example.com/login")
browser_type("[name='email']", "user@example.com")
browser_type("[name='password']", "secret")
browser_click("[type='submit']")

Multi-tab Research

> Open GitHub trending and compare top 2 repos

browser_navigate("https://github.com/trending")
browser_get_element_text("article.Box-row:first-child h2 a")
browser_click("article.Box-row:first-child h2 a")
browser_screenshot()

Semantic Targeting

> Use stable refs instead of brittle selectors

browser_get_snapshot()
browser_click({ "ref": "e42" })
browser_type({ "label": "Email address" }, "user@example.com")

Session Isolation

Each MCP client gets its own isolated browser window. The side panel shows active sessions and allows manual tab routing.

Development

cd daemon && bun run --watch start   # Daemon with hot reload
cd server && bun run --watch start   # Server with hot reload
cd client && bun run --watch start   # Client with hot reload

License

MIT - see LICENSE