lilbee

lilbee is a single-executable local AI stack. It runs and manages local models (chat, embedding, vision, rerank) across available GPUs, and uses them as a conversational search engine over your files, notes, code and crawled web pages, with answers that cite the exact file and line. Key capabilities described in the README: - Model management: built on llama.cpp, with its own model manager and multi-GPU fleet. Browse Hugging Face, download models, assign roles, and run them on Metal, Vulkan, CUDA or ROCm. It does not require Ollama or LM Studio, but can point at an existing setup. - Interfaces: terminal UI, CLI, Model Context Protocol (MCP) server for coding agents, HTTP/REST API, and a Python library. An Obsidian community plugin exposes the same engine inside a vault. - Ingestion and search: indexes local files and code, and includes a web crawler that adds sites to the library. Answers are cited back to source locations. - Privacy: files, index, embeddings, questions and answers stay on the machine; no telemetry, no account, and no cloud calls unless a cloud model is configured. - Installation channels: standalone binaries plus Homebrew, AUR, Nix, Docker (GHCR), Flatpak, Snap, Scoop and npm launchers, as well as pip/uv for developers. Vendor-specific builds exist for NVIDIA CUDA, AMD ROCm, and pre-AVX2 CPUs (compat build). - Platforms: macOS, Linux and Windows. Linux builds bundle the Vulkan loader; GPU detection needs the system Vulkan loader installed. - Optional OCR for scanned PDFs and images via Tesseract or a GGUF vision model. The project is explicitly labeled active beta: releases on PyPI are pre-releases requiring `--pre`, and interfaces, command names and on-disk formats may change between betas.