ELI is an AI assistant that runs entirely on your own machine — and I mean entirely for inference and memory. It holds a real conversation, drives your computer, reads your screen and your documents, writes and fixes code, and quietly builds up a private picture of who you are the more you use it. Talk to it or type to it — your call. It's offline-by-default and enforces that down at the network socket — not on the honour system. Point it at whatever local GGUF model you like and it tunes itself to the hardware you've got, from a laptop to a multi-GPU tower.

v2.3.93 — CI green + docs/metrics accuracy. Fixes silent-swallow ratchet (172 ceiling); router gaps for MINIMIZE_WINDOW / FIX_FILE; live-refreshed docs, blueprints, and capability manifest (11,711 tests, 227 capabilities, ~190k LOC). Grab it from Releases.

v2.3.92 — cross-OS media + local screen intelligence. YouTube visible mpv playback; bundled-python yt-dlp discovery; grounded install offers for mpv/tesseract/ydotool per OS; local UI-grounding hooks (Phi-Ground/OS-Atlas/UGround/UI-TARS via GGUF in models/, no URLs); deep screen audit with memory/research fusion; Screen tab one-click settings. Ported to v3. Grab it from Releases.

v2.3.85 — redistributability + observability. Removed Linux-only TILE_WINDOWS override (uses cross-platform portable_app_control); international crisis resources; gated CLOSE_APP wmctrl; cross-platform file open in experimental tab; canonical SQLite paths in generated scripts. Silent passes: 646 → 172. Full private inventory: blueprints/SILENT_EXCEPTION_FULL_REPORT.md (gitignored). Public summary: docs/SILENT_EXCEPTION_AUDIT.md. Grab it from Releases.

v2.3.84 — observability hardening + release fix. Removed 469 silent except: pass handlers (646 → 172) across runtime, media, memory, perception, and GUI paths — failures now log at debug instead of vanishing. Full inventory: docs/SILENT_EXCEPTION_AUDIT.md. CI ratchet lowered to 177. Grab it from Releases.

v2.3.83 — cross-platform media foundation. Shared YouTube/mpv core, startup capability report, wpctl/pactl/amixer volume chain, ELI_BROWSER on all OSes, v3 effector decomposition for browser/volume helpers. eli/integrations/media/cross_platform.py (Linux MPRIS, macOS AppleScript, Windows media keys/URI, Android termux-open). Ported to v3 effectors + media_runtime. Grab it from Releases.

v2.3.80 — persona self-knowledge routing. “What do you know about yourself?” and depth follow-ups stay in CHAT with verified self-facts instead of repeating the SELF_REPORT runtime blurb. Grab it from Releases.

v2 is live software, not a polished product drop. ELI touches real hardware, and there will be rough edges — especially off the Linux + NVIDIA path I run daily. I'd rather get a bug report than a compliment: open an issue. ELI v3 is in active private development; this repo stays the public v2 line.

Contents

• What is ELI?
• Screenshots
• What it does
• Design principles
• Quick Start
• Choose your model
• Features
• Optional: a remote view over your own network
• Privacy
• Security
• Tested on & known limitations
• Under the hood
• Documentation
• Project status & contributing
• License · Contact

What is ELI?

ELI is a private AI assistant that runs entirely on your own computer. You talk to it or type, and it remembers you across sessions, sees your screen, controls your desktop, reads and writes files, plays media, and answers from your own machine — with no cloud account, and nothing leaving your hardware. Point your phone at it over your own Wi-Fi and it's there too.

Under that simple surface it's a local cognitive runtime, not a chatbot bolted onto someone else's API: a 12-stage reasoning pipeline, a fleet of specialist agents on a DAG orchestrator, layered memory (SQLite + a FAISS vector index + a knowledge graph), local voice and vision, and its own smart-home server — all behind a single typed or spoken interface with 227 capabilities. You bring your own GGUF model; ELI is model-, user-, and hardware-agnostic by design.

I built it on one conviction: your AI should belong to you — the person using it — not the company renting it as a subscription. That isn't a slogan on a badge; it's the constraint every architectural decision in here has to answer to:

• Everything is local by default. Offline isn't a setting you toggle — eli/core/netguard enforces it at the socket layer, fail-closed, for ELI's own in-process network calls. (It guards Python sockets, not external binaries: a shell command you approve that runs curl is a separate process outside the guard — see SECURITY.md.) Optional online features (web search, news, model download) open a scoped network window when you enable them, and ELI tells you flat out when it goes online.
• Your model of you stays on your machine. ELI learns your patterns and preferences and writes them to your disk, never a data centre.
• It improves for you, not for a vendor. Self-training means ELI gets better at serving you — not at serving someone's next corporate model update.
• Privacy isn't a feature checkbox. It's the founding constraint of the whole design.

There's an underdog logic to it, too. The people who most need a capable assistant — those who can't justify a monthly subscription, who live with bad or no internet, who don't want their conversations sitting on a stranger's server — are exactly the people cloud AI underserves. And because the model is yours and swappable, ELI's ceiling isn't fixed — it rises with the open-model frontier and the hardware you give it. The quality is a dial you control, not a subscription tier someone sells you back.

Screenshots

The desktop app and the phone/web dashboard, running fully local:

Desktop — chat driving the OS (Spotify, volume, news)	Web dashboard — cognition mesh: 15 agents on the DAG orchestrator

Web chat — ELI introducing itself	Web overview — vitals, tasks, media & devices

More in .github/screenshots/ — including the phone-connect page (QR pairing, LAN-only).

What it does

You talk to it or type — it's not fussy about exact wording. The sort of thing I use it for day to day:

Asking "what can you do?" lists the full surface.

Design principles

What sets ELI apart isn't cosmetic — it's baked into the architecture. Each of these is a real mechanism in the code, not a line I liked the sound of:

1. Local and offline-first. A process-wide network guard fail-closes at the socket layer: with networking disabled, no outbound connection can be made, even by a component that tries.
2. Model-agnostic. No model name or size is hardcoded on the inference path. ELI loads any local GGUF, detects its chat template, and sizes context to the model's real n_ctx_train.
3. Grounded introspection. Asked about its own state, ELI reports from live runtime evidence — actual database row counts, the loaded model, the active pipeline — rather than generating a plausible answer. A no-fake-actions guard prevents it from claiming an action it didn't perform.
4. Self-maintaining. It logs its own failures and can generate, syntax-check, apply, and automatically roll back patches to its own source. A LoRA/QLoRA pipeline (PyTorch/PEFT) can fine-tune the model on your own conversation history.
5. Embodied. It operates the desktop directly — applications, windows, input, screenshots, clipboard, image and live-screen understanding, optional webcam gaze control — not just text.
6. User-aware. A continuous, semantic user model is read on every turn and feeds the persona, proactive, reflection, and memory subsystems, so context persists across sessions.

Quick Start

Installers (recommended) — Windows · macOS · Linux

Grab the latest from GitHub Releases. Every release is built in CI and launch-tested on all three platforms before it can publish — the pipeline boots each bundle and verifies the full runtime stack, bundled voices and data paths, or the release is rejected.

First launch walks you through everything:

1. GPU — ELI detects your hardware and offers acceleration: NVIDIA → CUDA, AMD / Intel Arc → Vulkan (a one-time download that is load-verified before it ever activates; CPU inference always works regardless). Apple GPUs need nothing.
2. Model — no multi-gigabyte weights ship inside the installers; ELI offers the recommended starter model sized to your VRAM with a progress bar. Add or switch to any local GGUF any time in the app.
3. Talk — voice works out of the box (Piper voices are bundled); type or speak.

Your data lives in one per-user folder (%LOCALAPPDATA%\ELI_v2, ~/Library/Application Support/ELI_v2, ~/.local/share/ELI_v2) and survives upgrades — settings, memory and models carry over like a save-game. Want a clean slate? Tick Fresh install in the Windows setup, or run ELI --fresh-start. GPU later? ELI --install-gpu-pack (add --vulkan for Intel Arc). Verify any download against SHA256SUMS.txt on the release.

# Without gh: download assets manually, then:
python3 scripts/restore_github_asset_files.py --from-dir /path/to/downloaded/assets

From source (developers)

Linux / macOS — install.sh gives you a system report → a plan → installs the right CPU/GPU build → offers to download a model sized to your hardware:

git clone https://github.com/ShadowESC95/ELI_v2.0.git
cd ELI_v2.0
bash install.sh                 # interactive: report → plan → install → pick model(s)
./scripts/eli_launch.sh         # launch the desktop app (first run shows a quick setup)

Flags: --yes (no prompts) · --install-cuda (auto-install CUDA toolkit) · --cpu-only · --model=qwen2.5-7b / --no-model.

Windows (double-click install.bat, or PowerShell):

.\install.bat            # CUDA + frozen lock + GPU verify
.\eli.bat                # launch

See docs/FIRST_RUN.md for clone vs portable paths.

Choose your model

ELI is not locked to Qwen, Mistral, or Phi. The installer catalog is a convenience menu only — inference loads any GGUF you place under models/ (recursive scan via discover_models()). That includes Kimi, Nemotron, Europa, Llama, Gemma, DeepSeek, MoE quant builds, and future families, as long as:

• The file is a chat/instruct GGUF (not an embedder-only weight — those live under models/embeddings/ and are auto-excluded from the chat picker).
• llama-cpp-python on your hardware can load it — ELI's smart-fit scales layers, context, and batch to the VRAM/RAM you have.
• For best reply quality, the GGUF should carry an embedded tokenizer.chat_template (most modern instruct quants do). ELI reads that metadata first, then falls back to filename heuristics (Qwen/ChatML, Llama-3, Mistral [INST], Phi, Gemma), then a generic prompt format.

Vision models need the matching mmproj GGUF alongside the main weights. Ollama is also supported as an alternate backend if you point ELI at a local Ollama instance.

To use a model: drop YourModel-Q4_K_M.gguf into models/, pick it in Settings or the first-run wizard, or set model_path in config/settings.json. The installer can also fetch one sized to your hardware:

python -m eli.core.model_download --choose   # multi-select menu — pick ANY number
python -m eli.core.model_download --auto     # one best-fit for your VRAM

The "needs" figures are for running each model at its most efficient. For what it's worth, I run Qwen3.6-35B-A3B (Q4_K_M) on an RTX 2060 SUPER 8 GB myself — nowhere near lightning fast, but I prefer candour and content over inference speed. Not that I currently have the choice.

The tiny embedder (memory/RAG) installs automatically with install.sh unless you pass --no-model. Vision and custom voice packs are optional extras. Fine-tune your own model — see docs/TRAINING_YOUR_OWN_MODEL.md.

Features

Conversation and reasoning. Five reasoning modes — Quick, Normal, Advanced, Research, Expert. Quick uses a single-pass reasoning algorithm at light orchestrator depth; Normal through Expert use genuinely multi-pass strategies (self-consistency sampling, tree-of-thoughts branches, draft → critique). Every CHAT mode — including Quick — runs the gradient orchestrator and 15-agent bus at a depth scaled to the mode; Quick is faster, not a bypass. The mode auto-selects by question depth, and when supporting evidence is weak ELI deepens on its own — re-gathering harder and escalating a tier before answering (with optional background deepening after a Quick reply). Underneath, retrieval combines HyDE query expansion (when eligible), vector + full-text + knowledge-graph search, re-rank, and synthesis — inside the 12-stage cognition pipeline (S01–S12 in eli/kernel/pipeline_trace.py).

Memory that persists. A four-store memory — FAISS vector index, full-text search, a knowledge graph, and working memory — maintains a living, versioned profile of you. Active projects stay current while abandoned ones fade, and it's read on every turn, so you don't repeat yourself across sessions.

Operates your computer. Open, close, focus, tile, minimise, or maximise apps and windows; switch workspaces; open URLs and your IDE; control volume; type text; move and click the mouse. App launch is backed by a live index of your machine's own executables — and if an app isn't installed, ELI offers to install it (real apt / snap / flatpak, on your confirmation).

Voice, hands-free. Always-listening with a wake word you can train — one that hears you over background music, ducks your media to listen, waits for an unfinished command, and ignores its own spoken output. Microphone auto-resolve (v2.3.44) probes ranked inputs on every OS — USB, Bluetooth/AirPods, and headset Chat endpoints before built-in/webcam/virtual routes — with live subprocess probes at native sample rates so silent false positives are rejected; Linux additionally pins live PipeWire/Pulse sources when needed. Run python -m eli.tools.mic_diag if you need to see what ELI picked. Dictation, transcription, and a Piper TTS voice included; a "train my voice" session learns your pitch, energy, and tone so delivery adapts to how you sound. Built-in character voices and a 166-voice/45-language downloadable library, all free and offline. An optional neural voice engine (Coqui XTTS-v2, pip install -e ".[natural]") adds human-like speech and lets you clone any voice from a short recording — drag an audio/video clip onto the drop zone in Settings ▸ Voice, or say "create a voice called X from clip.mp4"; ~1.8GB one-time model download, non-commercial use (Coqui CPML licence).

Vision, screen, and gaze. Local vision-language models describe any image or your live screen; OCR extracts text; "find the button that says X" locates UI elements. Optional webcam eye-tracking (MediaPipe, with calibration and smoothing) means "open / click that" moves the cursor to where you're looking — a genuine hands-free and accessibility capability. All local, no cloud vision APIs.

Image generation. A from-scratch procedural renderer with 10+ scene types (landscape, space, city, poster, emblem, abstract, product, …) — no model required. Plus optional SSD-1B diffusion with VRAM hot-swap, and matplotlib plotting from your data.

Documents and files. Create, read, summarise, and analyse files (CSVs, PDFs, images — single or whole folders); convert any document to PDF, .docx, .odt, .rtf, HTML, Markdown, .tex, EPUB, or .txt (pandoc + LibreOffice fallback). Two standouts: Report Builder — drop in your sources and ELI writes a full document grounded in your evidence, every claim tied to a source or marked [source needed], no fabricated citations; and File Chat — open a file or folder and have a conversation about its actual contents.

Coding agent. Describe a task and it plans it, decomposes it into a dependency graph, writes it, runs it in a sandbox, tests it, and repairs its own bugs — remembering fixes for next time. Plus examine-and-fix on your existing files (tiered scan → offer → verified, auto-reverting patch), project scaffolding, diffs, and a built-in Sim-IDE. The IDE ships with ELI's own native code editor (line numbers, Python syntax highlighting, current-line highlight, auto-indent); on a PyQt6 install with PyQt6-QScintilla present it upgrades to QScintilla automatically.

Scheduling and automation. Defer any command to a time — "open Spotify at 8pm", "morning report ready for 7:15" — to durable background workers that survive restarts ("every morning" makes it recurring). Catch-up jobs wait until you finish a live conversation before running, so a missed overnight task cannot hijack an active chat stream. Chain several commands in one sentence. Alarms, timers, pomodoro included.

Proactive and self-maintaining. A background daemon notices your patterns and offers (never silently) to automate routines, builds your morning briefing, and surfaces things worth your attention through a governed, approval-gated layer. ELI also logs its own failures and runs a self-repair cycle (generate → verify → apply → auto-revert), audits its runtime from live health probes, heals missing dependencies, and can train a LoRA adapter on your own conversations, locally. Nothing destructive runs unattended.

Make it yours. Swap the model (any local GGUF). Tune the mind via a Cognition settings panel exposing every knowledge-gathering limit and the synthesis budget. Extend it with a real plugin system. Teach it routines it proposes. Control the boundary with a single network toggle.

Optional: a remote view over your own network

The desktop app is the main event, but ELI also has a small built-in home server (local FastAPI) so you can pull it up in any phone/tablet browser on your own home network as a second screen. The AI still runs entirely on your computer — the browser is just a window onto it; nothing runs on the device and nothing reaches the internet.

Installer builds ship it ready to go: the ELI Server shortcut (Windows / Linux menu entry) opens it with a visible console showing the phone-connect URL and QR — or run ELI --server. From source:

./scripts/eli_serve.sh             # this computer only  → http://127.0.0.1:8081/
./scripts/eli_serve.sh --lan       # your home network   → prints a token-protected URL

The web app is an installable PWA with ten tabs: live dashboard, chat (streaming, markdown, multi-session, local voice in/out), a searchable command catalogue, smart-home control (ELI's own MQTT device server — device discovery, rooms, scenes, automations, no Home Assistant, no vendor cloud), measured system telemetry, shared research corpora (grounded, cited, attributed Q&A over your own documents), a tamper-evident HMAC-chained audit trail, an admin console with roles (admin / member / viewer), live settings, and QR-code pairing.

Server security in brief — full detail in docs/SERVER_AND_WEB_APP.md:

• The auth gate fails closed. No ELI_API_TOKEN and no explicit opt-out means every action endpoint returns 401 — even under a raw uvicorn / Docker / systemd launch that never runs main().
• Any non-loopback bind requires a token. Expose beyond loopback with no token and ELI auto-generates one and prints it rather than ever serving unauthenticated. Loopback stays zero-friction (tokenless only for a genuine loopback bind; ELI_API_ALLOW_TOKENLESS=0 to require one anyway).
• Roles are opt-in and attribution is authenticated. Viewers are read-only, members are 403'd from the admin console, and a member can't claim to be someone else in the audit trail. RBAC user tokens are stored as SHA-256 hashes; the single-operator LAN pairing token is a plaintext local file (config/api_token, 0600 on Unix) compared with a constant-time check.
• Loopback is trusted by default. A process on the same machine reaching 127.0.0.1 is treated as the local operator (tokenless admin) even when the server is bound for LAN access — deliberate for single-user desktops, a footgun behind a same-host reverse proxy or on shared hosts. Opt out with ELI_LOOPBACK_ADMIN=0. Details in SECURITY.md.
• Research ingest is sandboxed to a configured root with file-count / byte caps — a client can never make the server read arbitrary host files.
• Keep it on your own LAN. Don't port-forward it to the public internet.

Privacy

Nothing leaves your computer unless you ask for something online (news, a search, a model download) — and ELI tells you flat out when it does. No accounts, no telemetry, no subscription, no asterisks. A fresh install knows nothing about you until you talk to it, and you can wipe your data whenever you want.

Security

ELI is a host-control agent — treat it like running a capable automation tool with an LLM attached, not like a sandboxed chat app. Defence-in-depth, all local:

Full detail: SECURITY.md. Found a security issue? Report it privately per SECURITY.md, not in a public issue.

Tested on & known limitations

Last updated 2026-09-08 (v2.3.93).

I'd rather tell you exactly what I've run than pretend it's flawless everywhere.

What I've actually run, end to end: Linux (x86_64) with an NVIDIA GPU — install, first-run, the full test suite (11,711 tests collected, 11,700+ passing), voice (USB/Bluetooth headset auto-resolve verified on Linux), vision, the server, the AppImage with the CUDA GPU pack, all of it. The Windows installer has also been field-run on real hardware — install, first-boot GPU + model flow, voice and chat. On top of that, every release is launch-tested in CI on Windows, macOS and Linux: the pipeline boots each built bundle and verifies the runtime stack, bundled voices and data paths before it may publish.

What the code handles but hasn't had a human run on real hardware yet: macOS (CI-tested every release, not yet human-confirmed) and AMD GPUs (the Vulkan pack is CI-built and load-verified at install; real-card reports welcome). The installers, per-OS requirements, cross-platform paths, GPU detection (NVIDIA / AMD / Apple), and the OS-abstraction layer are all there and I've read every line — but "correct in the code" isn't the same as "confirmed on the metal," and I won't claim it is. If you run ELI on one of these, I'd genuinely love the output (working or broken) so I can close the gap for everyone.

The honest edges:

• Desktop control varies by OS. The core — chat, memory, the model, the server, vision — is portable. The part that drives your actual desktop (window control, some audio, screenshots) is built on Linux and routed through a cross-platform layer; it's the piece most likely to be rough on Windows/macOS on first run.
• macOS needs permissions. The first time ELI takes a screenshot or moves the mouse, macOS asks for Screen-Recording and Accessibility access (System Settings → Privacy). Grant and retry.
• AMD voice is CPU-only. The speech-to-text engine (CTranslate2) has no ROCm support, so on an AMD GPU it stays on the CPU — it works, just not accelerated. Everything else uses your AMD GPU.
• Windows secret-file permissions are weaker. ELI locks its token/key files to 0600 on Unix; Windows ignores that bit, so those files lean on your user profile being private (which it is by default under %APPDATA%). A proper Windows ACL is on the list.
• Big models can be slow to load or run out of memory. A large model takes a minute on first use, and if VRAM/RAM is tight the server can drop mid-reply. Running the server on its own (not alongside the desktop app) loads the model once and avoids most of that.
• Android/Termux is headless — server + CLI only, no desktop GUI.

None of these are hidden gotchas — they're the honest edges of a local, embodied assistant that touches real hardware. I'd rather you know them going in.

Under the hood

python3 -m venv .venv && . .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel
python -m pip install -r requirements-full.txt
python -m pip install -e .[full]
python -m eli            # GUI   ·   python -m eli --headless   # terminal REPL

bash scripts/build_v2_release.sh                    # v2 portable Linux (download-and-run)
bash scripts/build_v2_release.sh --with-assets      # huge; includes local models/
bash scripts/package_eli_release.sh                 # wheel/sdist
bash scripts/package_desktop_app.sh                 # portable Linux desktop package
bash build_packages.sh wheel deb appimage macos windows

Documentation

• Server & Web App — self-hosted FastAPI server + phone/web UI
• Train your own model — A-to-Z LoRA/QLoRA into an ELI GGUF
• Cross-platform coverage — capability × platform matrix
• Model runtime policy — how ctx/layers/batch are sized
• First run — clone vs portable paths
• Release pipeline — how the installers are built, launch-tested and published; GPU packs

Project status & contributing

ELI v2.0 is here to use. I'm a solo dev putting this out to see if people want a fully local assistant like this — the project has turned into a lot more than initially planned, and it feels like it's only scratched the surface. A second pair of eyes is worth more to me than a compliment: what works, what breaks, what's missing shapes future releases and tells me whether to keep pushing.

• Found a bug or have an idea? Open an issue — bugs, ideas, or a plain "I tried it and…" are all invaluable.
• Want to contribute code? Pull requests are welcome — please read CONTRIBUTING.md first. Because ELI is source-available, contributions include a short inbound license grant so the whole project stays under one consistent license.
• Security issue? See SECURITY.md — report it privately, not publicly.
• Want to support development? ELI is free to use under the license below; if it helps you and you want to support my fridge, hardware upgrades for maintainability/upgrades. ko-fi.com/shadowesc95 is optional and keeps me shipping, and to be honest, Enjoy, use, and report!

License

ELI v2.0 is source-available, not open-source, under the PolyForm Internal Use License 1.0.0 — © 2026 Jason Fitzgibbon Bridgeman.

Why PolyForm? I want anyone to be able to read, run, and modify ELI on their own machine for free — and I want commercial and distribution rights to stay with me while v2 is out in the open and v3 is in development. PolyForm Internal Use is the cleanest off-the-shelf license that does exactly that, without pretending to be open source when it isn't.

Forks for public re-publishing violate the license — contribute improvements back here instead of publishing your own copy. Seeing ELI redistributed, hosted, or sold somewhere? Please report it — with a link — to jaybridgeman0095@gmail.com.

Custom work? Personalised ELI builds, themes, training, and commercial licensing are available by email — separate from the free PolyForm grant above. Provided "as is", without warranty.

Third-party components: see THIRD_PARTY_NOTICES.md. Model/voice assets: see models/MODEL_LICENSES.md (also bundled with GitHub Release model packs).

Special Thanks

Node815 – For tireless testing on Arch Linux and hybrid GPU setups. Their detailed logs and persistence led directly to the cross-platform fixes in v2.1.19–v2.1.23, including:

• SQLite WAL fallback for NTFS/exFAT filesystems
• Qt xcb library bundling in the AppImage
• GPU detection for hybrid Intel+NVIDIA systems
• Ollama selector improvements

Thank you for making ELI better for everyone.

Contact

Questions, feedback, or interested in a license/services beyond the terms above?

• Email: jaybridgeman0095@gmail.com
• GitHub: @ShadowESC95 · open an issue