NavAgent

NavAgent — Ultra-light MCP browser navigation NavAgent lets your AI assistant (Claude, Cursor, Windsurf, Zed, OpenClaw, or any MCP-compatible client) control your browser using your real Chrome session. Your cookies, your logins, no cloud proxy. The AI sees a compact numbered list of clickable elements instead of screenshots or accessibility trees. This makes interactions fast, token-efficient, and anti-bot-proof. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ WHAT YOU NEED ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ NavAgent has two components that work together: 1. This Chrome extension (you're here) 2. The MCP server: npx navagent-mcp (npm package) Both are required. The extension scans web pages, the MCP server connects your AI client to the extension via WebSocket. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ SETUP (2 minutes) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 1: Install this extension. Step 2: Add the MCP server to your AI client config: Claude Code — add to .mcp.json: {"mcpServers":{"navagent":{"command":"npx","args":["-y","navagent-mcp"]}}} Claude Desktop — edit claude_desktop_config.json (same format as above) Cursor — edit ~/.cursor/mcp.json (same format) Windsurf — edit ~/.windsurf/mcp.json (same format) OpenClaw — add to openclaw.json under mcp.servers Any MCP client — stdio transport, command: npx, args: ["-y", "navagent-mcp"] Step 3: Open Chrome, start your AI client, and ask it: "go to https://example.com and scan the page" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ HOW IT WORKS ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ When the AI calls browse_scan, the extension scans the page and returns something like: 1. My Account [link] 2. Cart (0) [link] 3. Search [input] 4. Electronics [link] 5. Books [link] The AI reads this list, decides what to do (e.g. browse_click(5)), and the extension performs the action. No screenshots, no heavy parsing — just a numbered list. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 12 TOOLS AVAILABLE ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ browse_scan — Scan the page (zones or flat list of clickable elements) browse_zone — Drill into a zone to see its elements browse_click — Click an element by number (auto-rescans) browse_type — Type into an input field browse_more — Show next batch of elements (pagination) browse_scroll — Physical scroll for lazy-loading / infinite scroll browse_read — Read visible page text (max 2000 chars) browse_extract — Full page content as structured markdown with pagination browse_goto — Navigate to a URL browse_back — Go back to previous page browse_list_tools — List WebMCP tools declared by the page browse_call_tool — Invoke a WebMCP tool ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ KEY FEATURES ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ • Token-efficient — numbered lists instead of screenshots or ARIA trees • Anti-bot-proof — uses Chrome native extension messaging, not CDP. Undetectable by Cloudflare, Akamai, etc. • Works on any site — SPAs (hash routing), shadow DOM (forced open), same-origin iframes, contenteditable editors (DraftJS, ProseMirror) • Real browser session — your cookies, your logins, no re-authentication needed • Zone detection — landmarks (nav, header, footer, dialog) for structured navigation on complex pages • Markdown extraction — full page content as clean markdown with pagination • WebMCP support — discover and invoke tools declared by pages via navigator.modelContext ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ TESTED ON ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Reddit, LinkedIn, YouTube, X.com, OVH Manager, and more. Shadow DOM, iframes, rich text editors — all work. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ PRIVACY & SECURITY ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ • WebSocket listens on localhost only (127.0.0.1) — no external connections • Zero telemetry — no data sent anywhere, ever • Minimal permissions: activeTab, scripting, storage, alarms • Fully open source: https://github.com/DimitriBouriez/navagent-mcp ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ LINKS ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ GitHub: https://github.com/DimitriBouriez/navagent-mcp npm: https://www.npmjs.com/package/navagent-mcp Author: Dimitri Bouriez License: MIT