Local AI apps for Intel OpenVINO. Nothing comes in, nothing goes out.
Monad is a family of desktop apps for Linux that run AI models entirely on your own computer,
on the CPU, the Intel GPU or the NPU.
No cloud services, no accounts, no telemetry: your voice, your
documents and your questions never leave the machine.
The apps · Install · Why Monad · Private by design · Requirements · Troubleshooting · Contributing
Note
Why "Monad"? For Leibniz, a monad is a self-sufficient unit with no windows through which anything could come in or go out. Monad apps work the same way: the AI comes to your computer and everything it does stays there. The symbol ⊙, a circle around a point, is the ancient sign of the monad, and it signs every app of the family.
Each app has its own repository, its own releases and its own color, and they all share the same foundations: OpenVINO, one models folder, one design.
| App | What it does | |
|---|---|---|
| OVMS Manager the models hub · install first |
A graphical front end for OpenVINO Model Server. Browse the ready-to-use OpenVINO models on Hugging Face, download them, serve them through an OpenAI-compatible API (for example to chat in Open WebUI), and convert other models with Optimum. | |
| Lectern lecture notebook |
A lecture notebook with live speech-to-text. It records from the microphone and transcribes in real time with Whisper, on the NPU (about 6 times faster than real time on a Core Ultra 7), the GPU or the CPU, organizing the text by course, lecture and category. It also installs lectern-transcribe, which transcribes every audio or video file in a folder. |
More apps, all based on AI and built the same way, will follow.
Monad apps are installed one by one, each from its own repository. Install them in this order:
Install what the apps need from your distribution (see System packages for the exact command). The installers also check for anything missing and tell you what to run.
OVMS Manager is the hub of the family: it finds and
downloads the models into ~/models, the folder every Monad app shares, and serves them to other apps.
Install it first.
git clone https://github.com/VarCode-ai/ovms-manager.git
./ovms-manager/install.shOr, from its latest release, download
OVMS-Manager-<version>-x86_64.AppImage, or OVMS-Manager-<version>.tar.gz to unpack and install with ./install.sh.
Lectern uses the Whisper models found in ~/models.
git clone https://github.com/VarCode-ai/lectern.git
./lectern/install.shOr, from its latest release, download
Lectern-<version>-x86_64.AppImage, or Lectern-<version>.tar.gz to unpack and install with ./install.sh.
Lectern also works on its own: lectern-transcribe can download its Whisper model the first time you use
it. Installing OVMS Manager first gives you one place to browse, download and remove every model.
- Open OVMS Manager → Explore Hugging Face.
- For chat, pick an
INT4model that fits your memory (for exampleQwen3-8B-int4-ov) and press Download with OVMS. - For Lectern, choose the Speech-to-Text type and download
whisper-large-v3-turbo-int8-ov.
Then open Lectern: the model is already there. On the NPU, the first time you use a model Lectern compiles it for a few minutes and saves the result; after that it starts in a second or two.
Every app offers the same ways in, and they behave the same everywhere:
- AppImage: one file with its own Python, Qt and OpenVINO. It installs nothing, needs no root, and runs
on any x86-64 distribution with glibc 2.34 or newer. Verify every download with the
SHA256SUMSof the same release. - Installer (
./install.shin each repository): works only in your home folder and never asks forsudo. It creates a private Python environment, adds the app to the application menu and a command to~/.local/bin. It is in the clone of the repository and in the<App>-<version>.tar.gzarchive of each release. Once done you can delete the folder.--uninstallremoves the app and keeps your models, notes and settings;--devruns the files of the clone;PYTHON=python3.12picks another Python.
When a new version of an app is published, that app shows a banner with a link to its release notes. To update, download the new AppImage, or pull the app's repository and run its installer again:
git -C ovms-manager pull && ./ovms-manager/install.sh
git -C lectern pull && ./lectern/install.shYour models, notes and settings are kept.
Tip
Did you install the apps from this repository when it was called ai-app-local-openvino? Clone the two
app repositories as shown above and run their installers: they update the apps in place, with your
models, notes and settings.
Recent Intel processors include an NPU (the AI Boost in Core Ultra chips) and a capable integrated GPU, and OpenVINO can run modern AI models on both. Using them, though, still means working in the terminal: installing drivers and runtimes, converting models, and writing code against the OpenVINO APIs.
Monad makes the NPU and OpenVINO easy to use from a graphical interface: you install an app, pick a model and an accelerator, and start working. Every app is built on the same architecture:
- OpenVINO runs every model locally, on the CPU, the GPU or the NPU.
- Shared models: all apps use one models folder (
~/models), so a model downloaded once is available to every app. - OVMS Manager is the common hub to find, download, convert and serve models.
- The NPU works out of the box: OVMS Manager's container image carries the NPU driver and compiler, and Lectern brings the same driver when your distribution's one has none. On a Core Ultra 7 355, Lectern transcribes a lecture about 6 times faster than real time on the NPU.
- Every app is a PySide6 desktop app with a per-user installer that needs no root access, built on the
shared UI kit in
common/.
AI assistants usually work by sending your voice, your documents and your questions to someone else's servers. Monad apps take the opposite approach: the AI comes to your computer, your data does not go anywhere.
- Everything runs locally. Lectern transcribes your microphone and files on your CPU, GPU or NPU. The models served by OVMS Manager answer from your own machine. Recordings, transcripts, prompts and answers never leave your computer.
- No accounts, no API keys, no subscriptions. There is nothing to sign up for, and no usage limit other than your hardware.
- No telemetry. The apps collect no usage statistics, analytics or crash reports. They also switch off
the optional usage statistics of the libraries they run: Hugging Face downloads
(
HF_HUB_DISABLE_TELEMETRY,DO_NOT_TRACK) and OpenVINO's model conversion in the Converter. - Works offline. Once the models are downloaded, transcription and inference need no connection at all.
- Your data stays yours. Notes are plain text files in folders you choose: no database, no proprietary format, no lock-in. Uninstalling an app never deletes your notes, models or settings.
- Recordings are local too. If you choose to save the audio of a lecture, it is a
.wavfile in the lecture folder, nowhere else.
The apps connect to the internet only in these cases. Each request reveals your IP address and a software name, as any HTTP request does; none of them carries your audio, your text or anything that identifies you.
| When | Where | What for | What is sent | Can you avoid it? |
|---|---|---|---|---|
| You open Explore Hugging Face (OVMS Manager) | huggingface.co | List of OpenVINO models | Your search text, if any | Yes: the catalog is cached for 6 hours and offline the cache is used |
| You download a model | huggingface.co | Model files | The model name | No: it is the download you asked for |
| You start the server or download for the first time | docker.io | The OpenVINO Model Server container image | The image name | No, but only once |
| First recording in Lectern | raw.githubusercontent.com | Silero voice detection model (1.3 MB) | Nothing | Yes: offline, Lectern uses a simpler volume-based detection |
| At most once a day | api.github.com | Is there a newer release of this app? | Nothing | Yes: Help → Check for updates automatically in Lectern, the command palette in OVMS Manager |
| You run the Converter | huggingface.co | The original model to convert | The model name | No: it is the conversion you asked for |
Buttons such as Open on Hugging Face or Open WebUI simply open your web browser.
- GPL-3.0-or-later. You are free to use the apps for any purpose, study how they work, change them and share them. Copyleft guarantees that every version based on this code stays free and open as well: these apps can never be turned into a closed product.
- Open from top to bottom. The apps depend only on open source software: OpenVINO and OpenVINO Model Server (Apache 2.0), Qt for Python (LGPL), Silero VAD (MIT), and the Inter and Newsreader fonts (OFL). The Whisper models are open-weight (MIT).
- Verifiable, not just promised. Every claim on this page can be checked in a few files:
common/ui_kit.py,UpdateChecker: the update check, shared by every app;ovms_manager.py,hf_getandNO_TELEMETRY_ENV: the catalog and the conversion;lectern.py,SileroVAD: the voice detection download;transcribe.py,resolve_model: the model download.
- Built in the open. Every change runs the public checks in GitHub Actions. Issues, translations and pull requests are welcome: see CONTRIBUTING.md.
Every Monad app shares the same kit, so they look and behave the same while each keeps its own color:
.
The brand itself (the name, the ⊙ symbol, the colors and the icon template) is described in
docs/brand.
- Light and dark theme that follow the system and switch live. You can force either one: Tools → Appearance in Lectern, the theme menu in the sidebar of OVMS Manager.
- Calm layout: warm neutral colors, a sidebar that recedes, few borders. Depth is used only for what floats, such as Lectern's recorder dock and the command palette.
- Command palette (Ctrl+K) in every app: every action, every lecture and every local model is one search away.
- Visible provenance: Lectern shows where the transcription runs (Local · GPU, updated if it falls back to the CPU), and a live waveform shows when it hears speech. OVMS Manager tells you whether a model fits in your RAM before you download it.
- Typography: Inter for the interface, with tabular figures for clocks and tables, and Newsreader as an optional serif for reading transcripts. Both fonts are bundled, so the apps look the same on every distribution.
- Accessibility: every text and focus color reaches the WCAG 2.2 AA contrast ratios in both themes
(checked by
tools/check_contrast.py), focus rings are visible, and Reduce motion turns off the animations. It follows GNOME's setting by default.
The apps are available in English and Italian. They follow the language of your system, and you can choose another one from the app: Tools → Interface language in Lectern, the 🌐 menu in the sidebar of OVMS Manager. The change applies the next time you start the app.
Translations are plain JSON files in each app's locale/ folder, so adding a language does not need any
tool: see the CONTRIBUTING guide of each app.
- Linux on x86-64 with glibc: Ubuntu 22.04+, Debian 12+, Fedora 39+, Arch Linux, openSUSE
Tumbleweed, Linux Mint, Pop!_OS and their derivatives. openSUSE Leap 15.6 works with the
python311package andPYTHON=python3.11(its defaultpython3is too old). Distributions based on musl (Alpine) are not supported, because PySide6 and OpenVINO only publish glibc wheels. - Python 3.10 or newer with the
venvmodule (only for the installers: the AppImages carry their own). - A desktop session (Wayland or X11).
- OVMS Manager also needs Podman or Docker to run the model server.
- Lectern also needs
ffmpegto read files other than.wav, and records best with PipeWire'spw-record.
Optional, for hardware acceleration:
- Intel GPU (integrated UHD/Iris Xe graphics or Arc). OVMS Manager needs nothing extra, since the container image ships its own GPU driver. Lectern runs Whisper directly on your system, so it needs the Intel compute runtime on the host (see the table below).
- Intel NPU (Core Ultra processors). Needs a recent kernel with the
intel_vpumodule and the NPU firmware, nothing else: OVMS Manager's container image ships its own NPU driver and compiler, and Lectern brings the same driver when the one of your system has no compiler (as in Fedora 44 and 45). Lectern also needs the Level Zero loader (oneapi-level-zeroon Fedora,libze1on Debian and Ubuntu).
The installers check for missing packages and print the right command for your system. To install everything up front, use the command for your distribution:
| Distribution | Base packages | Intel GPU runtime (for Lectern) |
|---|---|---|
| Ubuntu, Debian, Mint | sudo apt install python3 python3-venv ffmpeg podman pipewire-bin libxcb-cursor0 |
sudo apt install intel-opencl-icd |
| Fedora | sudo dnf install python3 ffmpeg-free podman pipewire-utils xcb-util-cursor |
sudo dnf install intel-compute-runtime |
| Fedora Atomic (Silverblue, Kinoite) | rpm-ostree install ffmpeg-free xcb-util-cursor (Python, Podman and PipeWire are preinstalled) |
rpm-ostree install intel-compute-runtime |
| Arch Linux, Manjaro, EndeavourOS | sudo pacman -S python ffmpeg podman pipewire xcb-util-cursor |
sudo pacman -S intel-compute-runtime |
| openSUSE | sudo zypper install python3 ffmpeg podman pipewire-tools libxcb-cursor0 |
sudo zypper install intel-opencl |
You can replace podman with Docker if you already use it. libxcb-cursor is only needed on X11
sessions, but installing it costs nothing.
These variables are shared by the apps; each app's README lists its own.
| Variable | Default | Used by | Meaning |
|---|---|---|---|
OPENVINO_MODELS_DIR |
~/models |
every app | Folder holding the OpenVINO models. |
OVMS_CONTAINER_RUNTIME |
auto (podman, then docker) |
OVMS Manager | Forces the container runtime. |
OVMS_EXTRA_RUN_ARGS |
empty | OVMS Manager | Extra options for every container, e.g. --group-add keep-groups. |
OPEN_WEBUI_URL |
http://localhost:3000 |
OVMS Manager | Address opened by the Open WebUI button. |
LECTERN_AUDIO |
pw-record when available |
Lectern | qt records the microphone through Qt instead of PipeWire. |
LECTERN_NPU_DRIVER |
auto |
Lectern | system never uses the Intel NPU driver bundled with Lectern, bundled always does. |
PYTHON |
python3 |
installers | Interpreter used to create the environments. |
The apps follow the XDG base directories: settings go in ~/.config, caches and logs in ~/.cache, and
the installed apps in ~/.local/share.
- The OVMS API has no authentication. By default the container publishes port 8000 on every network interface, so a containerized Open WebUI can reach it. On untrusted networks, enable Localhost only on the Server page or block the port with your firewall.
- Allow custom code (
--trust-remote-code) in the Converter runs Python code from the model repository on your computer. Enable it only for repositories you trust. - Models come from their authors on Hugging Face. Prefer the official
OpenVINO/repositories, which OVMS Manager marks with ✓. - To opt out of OpenVINO's usage statistics for every OpenVINO tool on your system, not only inside these
apps, run
opt_in_out --opt_outonce.
This is the home of the family: the guide you are reading, the brand, and the reference copy of the kit every app is built on. The apps themselves live in their own repositories.
monad/
├── common/ # the shared UI kit, copied into every app
│ ├── ui_kit.py # design tokens, themes, icons, command palette, update checker
│ └── fonts/ # Inter and Newsreader (SIL Open Font License)
├── docs/
│ ├── brand/ # the Monad logo, symbol, app icons and brand guide
│ ├── images/ # screenshots
│ └── troubleshooting.md # problems common to every app
├── tools/
│ ├── check_contrast.py # checks the WCAG contrast of the theme colors
│ └── sync-common.sh # copies the kit into the app repositories
└── .github/workflows/ # checks on every push
docs/troubleshooting.md covers the problems common to every app: installation, Qt failing to start, the NPU. Each app has its own guide too: OVMS Manager (containers, the GPU in Podman) and Lectern (the Intel GPU, audio).
Bug reports and pull requests are welcome. Report a problem with an app in that app's repository; the kit, the brand and the documentation of the family belong here. CONTRIBUTING.md explains how the kit is shared and how to build a new Monad app.
Copyright © 2026 VarCode-ai.
Monad and its apps are free software, released under the GNU General Public License v3.0 or later.
They build on third-party software that keeps its own license: OpenVINO, OpenVINO GenAI and OpenVINO
Model Server (Apache 2.0), Qt for Python / PySide6 (LGPL v3), Silero VAD (MIT), OpenAI Whisper models
(MIT), the Inter and Newsreader fonts (SIL Open Font License 1.1, see common/fonts/), and the models you
download, each covered by the license on its Hugging Face page.

