We built LHIC (Local Human Intent Controller) because we were frustrated by existing LLM-based browser agents. Tools like WebVoyager or ServiceNow's baseline agents are amazing, but they are: 1. Slow: Taking 20 to 40 seconds per simple form-fill step. 2. Expensive: Calling GPT or Claude for every single click. 3. Brittle: Failing the second a minor DOM class name shifts. LHIC introduces a "Fast Path / Slow Path" dual-route execution architecture: - Fast Path (Local, 0 LLM call): For standard web flows (logins, forms, search), LHIC matches user intents directly with pre-defined skills and resolved DOM attributes locally. It runs in under 35ms (median) with zero API call costs. - Slow Path (LLM, Fallback): Only activates for complex, open-ended tasks or when a high-risk action is flagged.
Key Technical Highlights: * Self-Healing Semantic Locators: We fuzzy-match DOM structural targets using SQLite historical skill memory, achieving an 80% success rate improvement over static CSS/XPath selectors when layouts change. * Enterprise Hardening: Features KMS Ed25519 signature checks (AWS/GCP/Vault), local AES-256-GCM database encryption, OTel APM export, and a custom Docker Seccomp profile to prevent Chrome escape. * VNC Screencast: Streams real-time page updates directly via CDP frame broadcasts for manual override. * MCP Integration: Standard Model Context Protocol stdio and HTTP gateway, ready to plug into Claude Desktop, Windsurf, or Cursor.
It's entirely open-source (MIT) and written in pure TypeScript with zero C++ compilation dependencies. We'd love to hear your thoughts on the architecture and hybrid execution concept!
Repo: [https://github.com/chengmatt416/LHIC](https://github.com/chengmatt416/LHIC)
Documentation: [https://lhic-docs.pages.dev](https://lhic-docs.pages.dev)
Comments URL: [https://news.ycombinator.com/item?id=48932029](https://news.ycombinator.com/item?id=48932029)
Points: 1