Hi HN — I built Runiq: a local “sovereign runtime” that lets LLM agents (Claude Desktop, local Llama, etc.) use hardened tools to interact with your OS, instead of executing arbitrary shell/Python.
Runiq implements Anthropic’s Model Context Protocol (MCP) and exposes a small set of controlled capabilities:
Stealth Chromium browser for scraping/automation (aimed to survive common bot checks)
File operations and local workflow tools
Human-in-the-loop security: native OS popups block risky actions (write/delete/etc.) until you approve
It ships as a single static Go binary (no venv/daemon). It’s localhost-only and auditable; if you pair it with a local model, nothing leaves your machine. There’s also an optional Python adapter to use the same tool interface with OpenAI/Groq/LangChain-style agents.
QaysHaji•2h ago
Runiq implements Anthropic’s Model Context Protocol (MCP) and exposes a small set of controlled capabilities:
Stealth Chromium browser for scraping/automation (aimed to survive common bot checks)
File operations and local workflow tools
Human-in-the-loop security: native OS popups block risky actions (write/delete/etc.) until you approve
It ships as a single static Go binary (no venv/daemon). It’s localhost-only and auditable; if you pair it with a local model, nothing leaves your machine. There’s also an optional Python adapter to use the same tool interface with OpenAI/Groq/LangChain-style agents.
Repo: https://github.com/qaysSE/runiq
I’d love feedback on the MCP implementation in Go, and on the security/UX tradeoffs for permission-gated agent actions.