About this project

ELI is a private, local-first AI assistant designed to run entirely on your own computer with no cloud dependencies. It combines a conversational interface with a cognitive runtime that includes a 12-stage reasoning pipeline, specialist agents on a DAG orchestrator, layered memory (SQLite, FAISS vector index, knowledge graph), local voice and vision capabilities, and a smart-home server. You can interact via typing or voice, and it can control your desktop, read your screen and documents, write code, and remember your preferences across sessions. The core philosophy is that your AI should belong to you, not a vendor. This is enforced at the architectural level: a network guard fail-closes at the socket layer, ensuring no outbound connections are made unless you explicitly enable optional online features like web search or news. Your data—conversation history, memory, and models—stays on your machine in a per-user folder that survives upgrades. ELI is model-agnostic. It loads any local GGUF chat/instruct model you provide, detects its chat template, and sizes context to the model's actual training window. It supports GPU acceleration (CUDA for NVIDIA, Vulkan for AMD/Intel Arc, Metal for Apple Silicon) and CPU inference. It also supports vision models via mmproj files and can use Ollama as an alternate backend. Key capabilities include: - **Computer control**: Open/close/focus apps, tile windows, adjust volume, type text, move mouse, click. - **Media playback**: Control Spotify and YouTube, get info on what's playing, skip ads. - **Screen & document understanding**: Summarize on-screen content, PDFs, images, and CSVs. - **Coding assistance**: Fix bugs, examine codebases, scaffold projects, write scripts. - **Memory & identity**: Remember facts, build a private profile of your preferences over time. - **Task automation**: Set alarms, timers, calendar, run scheduled jobs, learn routines. - **Voice interaction**: Wake word, dictation, text-to-speech with style presets and optional voice cloning. - **Self-maintenance**: Analyze failures, apply self-patches with rollback, and fine-tune itself via LoRA/QLoRA on your own conversations. Installation is straightforward with installers for Windows, macOS, and Linux (AppImage). First-run setup guides you through GPU detection, model selection (sized to your VRAM), and voice configuration. A web dashboard is included for remote access over your own network via QR pairing. The project is under active development (v2.4.37 as of the README) with a focus on stability and cross-platform compatibility. It's tested on Linux and NVIDIA primarily, with other platforms supported but possibly having rough edges. Contributors are encouraged to report bugs rather than compliments. The license is PolyForm Internal Use, and it supports funding via Ko-fi.