OpenCradle

OpenCradle (Cradle) ingests messages from Telegram, a web widget, or an HTTP API and triages them with local agents — no cloud calls, inference runs through node-llama-cpp.

A two-layer risk model (fast rule-based L1, LLM-classifier L2) decides what an agent can act on autonomously and what needs an operator to confirm, with a full audit log.

RAG runs on sqlite-vec with local embeddings over markdown, PDF, DOCX, and XLSX sources, and the whole stack ships as an Electron desktop app or a headless server daemon.