An all-in-one AI application that runs on your own computer. Chat with AI models, search your own documents, build agents, generate images — without sending any of it to a cloud service.
This repository is the public, open-source home of the app — it contains the source code, released under Apache-2.0. This page is the user guide: how to download, install and use the app itself.
New here? You are in the right place. Start with Getting started below. You do not need to be a developer, and you do not need to understand any of the technical terms on this page — every one of them is explained in plain language in the Glossary.
Already run local models? Skip all of that → the technical summary: pinned llama.cpp build, launch-flag handling, the RAG defaults, and what isn't implemented.
| I want to… | Go to |
|---|---|
| I already run llama.cpp / Ollama / LM Studio | Technical summary — skip the hand-holding |
| See what it can do | Feature tour |
| Download the app (new to GitHub?) | Downloading from GitHub |
| Install it on Windows | Windows installation guide |
| Install it on Linux | Linux installation guide |
| Install and operate it with an external AI agent | Agentic Support guide |
| Know what happens on first launch | First run |
| Fix a problem | FAQ & troubleshooting |
| Understand a word I don't know | Glossary |
| Know what leaves my computer | Privacy & your data |
| Connect a cloud or remote model | Cloud providers |
| Install curated apps that run beside it | External Apps |
| Update to a newer build | Updating |
| Run and maintain a node (exit codes, backups, downgrade, what phones home) | Operating a node |
| Send feedback or report a bug | Giving feedback |
| Download the latest build | Releases |
A conversation answered by a model running on the machine itself. Nothing in this exchange left the computer.
The app measures your hardware and tells you which models will actually run on it — instead of leaving you to guess.
Your own documents, searched locally and used to answer questions.
→ See every feature, explained with screenshots
| Minimum | Comfortable | |
|---|---|---|
| Operating system | Windows 10/11 (64-bit) or a current 64-bit Linux | Windows 11, or Linux with an up-to-date GPU driver |
| Memory (RAM) | 8 GB | 16 GB or more |
| Free disk space | 5 GB | 30 GB or more (models are large) |
| Graphics card | Not required — works on the processor alone | An NVIDIA, AMD or Intel GPU makes it much faster |
| Internet | Required for the first launch and downloads | — |
A graphics card is optional. Without one the app still works; answers just arrive more slowly.
Windows prerequisite: install the x64 ASP.NET Core Runtime 10.0.12 or a newer .NET 10 servicing patch. The Windows Portable ZIP does not bundle .NET. Linux remains self-contained and needs no system .NET installation.
Windows prerequisite: install the Microsoft Edge WebView2 Evergreen Runtime as well. The app's own window is built on it, and the app does not download it for you. Without it the window cannot open; starting the app with
--browseropens it in your web browser instead.
Windows and Linux, both x64. Both official downloads are portable Velopack applications and can update themselves. Windows ships as a Portable ZIP; Linux ships as an AppImage. There is no macOS or ARM build.
GitHub is built for programmers and the download page confuses almost everyone the first time. Here is exactly what to do.
1. Open the releases page: → Releases
2. <> Code button if you just want to run the app. It downloads this
repository's source code, not a ready-to-run build — there is nothing to double-click inside it.
That button is for developers who want to build the app themselves; everyone else wants the Releases
page instead. This is the single most common mistake.
3. Scroll to "Assets" and click it to expand. It is often collapsed behind a small ► triangle. You will see a list like this:
▼ Assets
XE-Local-AI-Engine-win-Portable.zip ✅ WINDOWS — download this
XE-Local-AI-Engine-<version>-linux.AppImage ✅ LINUX — download the .AppImage
CHECKSUMS.sha256 ✅ download this too
XE-Local-AI-Engine-<version>-delta.nupkg ❌ updater file; ignore
XE-Local-AI-Engine-<version>-full.nupkg ❌ updater file; ignore
releases.win.json / releases.linux.json ❌ updater feeds; ignore
RELEASE-MANIFEST.json / RELEASE.spdx.json ❌ release records; ignore
Source code (zip) ❌ ignore
Source code (tar.gz) ❌ ignore
4. Windows: click XE-Local-AI-Engine-win-Portable.zip. Linux: click the .AppImage. Download
CHECKSUMS.sha256 too, verify the platform artifact, then follow the matching installation guide. The .nupkg and
feed files are for the app's updater; you never open them manually.
5. If your browser blocks it (Edge/Chrome: "is not commonly downloaded"), open your downloads
list with Ctrl+J, click the ⋯ next to the file and choose Keep → Keep anyway. Same cause
as the Windows warning below: the build is unsigned.
Check you got the right file: Windows downloads XE-Local-AI-Engine-win-Portable.zip; Linux downloads an
.AppImage. If you downloaded GitHub's Source code archive, go back to point 2.
There is no installer and no Setup.exe. Windows users extract the Portable ZIP to a writable local directory;
Linux users mark the AppImage executable and run it.
More detail, checksum verification, and how to get emailed about new builds: Downloading from GitHub
Follow the Windows installation guide or Linux installation guide.
The app is not code-signed yet, so Windows shows a blue "Windows protected your PC" box.
The "Run anyway" button is hidden until you click "More info" first. Many people get stuck here and assume the app is broken — it isn't.
The installation guide has step-by-step screenshots of exactly what to click, plus a way to avoid the warning entirely.
Only "Don't run" is visible. "More info" is the small link — click it and "Run anyway" appears.
Everything below runs on your own machine unless you deliberately connect an outside service.
A summary — the feature tour shows each of these with screenshots.
The basics — what most people will use
- Chat with AI models running locally, with streaming answers — including sending an image to a model that can see
- Find and download models from Hugging Face without leaving the app
- Import a GGUF you already have — one file at a time, copied into the app's own models folder
- Hardware-fit advice — the app measures your RAM, VRAM and GPU, then recommends models that genuinely fit, including which quality/size trade-off ("quantization") to pick. (VRAM can't be read on AMD/Intel under Windows yet, so advice there is less precise.)
- Documents & knowledge bases — add your own files, or index a Git repository, and ask questions about them
- Its own app window — on Windows and Linux the app opens in a native window rather than a browser tab. On Windows, the first time you close it, it asks whether to exit or keep running in the tray
- Memory at a glance — the top bar shows how much RAM and graphics memory (VRAM) is in use and which models are loaded, and lets you eject one
- Updates you choose — pick Stable (released versions only), Preview (also release candidates) or Development (also daily test builds) on the About page. Details →
Going further — for people who want to build things
- Agents — configurable assistants with their own instructions, persona, tools, skills and memory
- Sub-agents — agents that can call other agents
- Scheduling — run a saved agent automatically on a timetable, with run history and cancellation
- Adaptive memory — agents remember useful facts across conversations, extracted by a local model. Candidates go through a best-effort scan for things that look like secrets first — pattern-based, so treat it as a safety net rather than a guarantee
- MCP servers — connect external tool servers to your agents
- Web search and page fetching — let a model search the web (DuckDuckGo, or your own SearXNG) and read public pages. Off by default; when on, the app asks you before each request is sent and shows you what came back before the model reads it
- Agentic Support — let a trusted same-machine external agent install and operate the node over
its loopback-only inbound MCP server. The restricted
delegatekey exposes 8 shared tools; theagentickey additionally exposes the administration tools listed in the shipped skill reference. See the Agentic Support guide before minting the higher-trust key - Custom tools — author an HTTP request or direct host-program launch for an agent. The node feature starts off and the built-in form initializes new tools as disabled. Every tool stays approval-wrapped: fixed tools may reuse an explicit session approval until edited, while parameterized tools ask on every call
- Skills — a local library of capabilities agents can load on demand
- Benchmarks — freeze one task of your own and run it against several models, with launch evidence, your own score, and an optional second model acting as judge
- Local model proxy — expose your local models over an OpenAI-compatible API so other programs on this machine can use them. Off until you generate a key, and loopback-only
Experimental — rough edges expected
- Development Mode — an agent works on a real Git repository of yours in an isolated copy, with a reviewed approval step before anything is written back. It ships enabled; you can switch it off under Node Settings → General → Features, and it acts only after you register a repository. Read the security boundary before you register one. Never point it at code you do not trust.
- Image generation — generate images locally, or edit one you generated or uploaded. Grouped as a preview feature in the app.
- Fine-tuning (Training) — build a training set with a local model, fine-tune a model on it, then score the result against the original on held-back samples. Linux with an NVIDIA graphics card only, and a run takes the whole GPU while it works.
- Graph Workflows — draw a workflow as a diagram (agent steps, tool calls, conditions, and a step that waits for your approval), then start runs of it and follow each one. A Chat workflow runs inside a conversation: pick it in Chat and it may ask you questions as it goes
- Audio transcription — local speech-to-text with whisper.cpp: upload a recording, or transcribe a microphone or shared-screen audio live in the browser. The audio is never stored
- Read answers aloud — text-to-speech through voices exposed by your browser and operating system. Availability, quality, and whether a system voice uses the network depend on that platform's speech implementation, which the app does not control.
- Optional cloud providers — Azure AI Foundry and Codex, if you want them. Entirely optional.
- Ollama — if you already run Ollama, the app can use its models. It will not install or manage Ollama for you.
When an agent or a tool runs commands, the app tries to fence it in: no access to the rest of your files, no network, and a cap on the CPU and memory it can use. How much of that your computer can actually provide differs from one machine to the next. You choose what the app does when the fence cannot be built in full. The app asks once, the first time you set it up, and an upgraded install starts on Low.
- Low — the app builds as much of the fence as your computer allows and runs the work anyway. Nothing is refused for a missing safeguard, and the Development page shows what was actually in place.
- High — a safeguard the work asks for becomes a requirement. If your computer cannot provide it, the work is refused with a message that names this setting, instead of running with a weaker fence. On Windows today, and on Linux without a per-user system service manager, High refuses every command-running feature, so the app recommends Low there.
You can change it at any time under Node Settings → Sandbox & isolation, which also lists which features High would refuse on your computer before you save. Switching from High to Low asks you to confirm, because it loosens the fence. Only you can change it in the app: an agent connected to the app cannot. Windows previews (the experimental isolation for Windows) are a separate switch in the same section, and High does not turn them on.
- No two-way voice chat — reading replies aloud and transcribing speech are separate features; the app does not hold a spoken conversation with you.
- No macOS build, no ARM build.
- Unsigned binaries — Windows and Linux releases can trigger trust warnings until certificate signing is added.
Verify
CHECKSUMS.sha256before running a download.
This is an early beta, built by one person in their spare time alongside a full-time job. Please read these before you start, so nothing comes as a surprise:
- The starter model is deliberately small. The app downloads a small model (IBM Granite 4.1 3B, ~2 GB) on first launch so chat and tools work out of the box. It is not representative of the quality this app can deliver — larger models answer noticeably better. Use the built-in advisor at Models → Recommendations to pick a real model for your hardware. How to do that →
- Windows will warn you on first launch, because the build is unsigned. What to click →
- Expect rough edges. This is early, actively-developed software.
- Your database is not fully encrypted. Sensitive fields are individually encrypted, but extracted document text is not. Details →
- Your password unlocks your data — don't lose both it and the recovery code. The key that encrypts your data is protected by your admin password, so the app asks for it every time it starts. Setup shows a one-time recovery code for a forgotten password. Lose both and the data cannot be recovered, by anyone. Details →
- Keep backups of anything you care about. Do not make this app the only place important data lives.
Stuck? Nothing is too basic a question — being stuck is itself useful feedback.
- Check the FAQ & troubleshooting — it covers the common problems.
- Still stuck? Open an issue or message me on Reddit.
When reporting a problem, the feedback guide explains what to include. Even "I opened it and didn't understand what to do next" is a genuinely valuable report.
- This is an early beta, built and maintained by one person.
- There is no required feedback report and no obligation to review or promote anything.
- The project is open source, licensed under Apache-2.0 — you're free to use, modify and redistribute it under that licence.
Thank you for trying it. Genuinely — honest impressions from people on hardware I do not own is the single most useful thing for this project right now.