Example applications and code samples built with the QVAC SDK — Tether's toolkit for running AI locally, privately, and on-device.
Each example here is a focused, self-contained app demonstrating one real-world use case for local AI: what becomes possible when intelligence runs on your own machine instead of someone else's cloud — no data leaving the device, no permission required.
The apps in this repository are prototypes and demonstrations.
- Provided as-is, with no support, no warranty, and no SLA.
- Not maintained as products — they may break, lag behind the SDK, or be removed.
- Not security-audited — do not use them in production or with real, sensitive data.
- They exist to illustrate use cases and teach.
See LICENSE for the full Apache 2.0 terms, including the disclaimer of warranty.
Every example lives in its own directory named qvac-{app-name} and is fully self-contained — its own code, its own setup steps, and its own README explaining what it does and why local AI matters for that use case.
| App | Description |
|---|---|
qvac-natural-language-to-sql |
Ask a banking database questions in plain English and watch a local AI (Qwen3 4B) write the SQL, then run it on-device against an in-memory SQLite bank. The schema and your question are the only things the model ever sees — no row data, no cloud, no API keys. |
qvac-story-image-generator |
Turn a child into the hero of a five-scene illustrated storybook, fully on-device. A local AI (Qwen3 4B) writes the captions while the child's own photo is composited into hand-drawn vector art right in the app. Only the story name and character ever reach the model — the photo never leaves the machine, and there is no image model to regenerate (and lose) the child's face. No cloud, no API keys. |
qvac-realtime-vision |
Point your webcam at the world and a local AI draws boxes around objects, tracks your hands and reads your gestures, and narrates the scene in one sentence, then turns the same on-device vision into two body-controlled mini-games. Detection and hands run on @qvac/onnx, narration on Qwen3-VL via @qvac/sdk. No cloud, no API keys, and offline after a one-time model download. |
qvac-gym-training |
Drop a short video of one set and a local AI (Qwen3.5-VL 4B) watches your lifting form and coaches it: what you do well, and what to fix. The app samples ten frames in the browser and the model reads them as one movement, all on-device via @qvac/sdk. The video never leaves the machine — only the frames reach the model, and there is no cloud and no API keys. |
qvac-smart-camera |
Point a camera at a doorway or driveway and a local AI understands what it sees: it boxes people, vehicles, animals and bags, and gives each person a risk verdict on-device (a resident coming home is normal; a masked person prowling a car at night is a high-risk alert with an alarm). Detection on @qvac/onnx (YOLOv10), risk verdict and description on Qwen3-VL via @qvac/sdk. A demonstration of on-device vision, not a real security system. No cloud, no API keys, no frame leaves the machine. |
qvac-voice-relay |
Enroll your own voice once, then type or speak a phrase and hear it played back in your voice, translated into another language. Speech recognition (Whisper), translation (Bergamot, pivoting through English), and reference-matched speech in 17 languages (Chatterbox) all run on @qvac/sdk. Consent-first enrollment, one-click erase, and your voice never leaves the device. |
qvac-creator-toolkit-demo |
Everything around a video except generating it, on-device: a local LLM writes the script (narration for one voice, or a dialogue), Supertonic 3 narrates it with a different voice per speaker, and Whisper builds timed .srt/.vtt subtitles from any video or audio file. Section labels like HOOK and OUTRO are stripped so they are never spoken, and it runs with the network off once the models are cached. No cloud, no API keys, no per-minute bill. |
qvac-desk-tidy-demo |
Point it at a messy folder and a local AI sorts the files by what they actually contain, not by extension: invoices, contracts, screenshots, photos, code, installers. Text is classified with EmbeddingGemma embeddings and images are described by Qwen3-VL then scored, both via @qvac/sdk. It shows the plan with a reason per file before moving anything, one click undoes a whole run, and a tray mode can keep a folder tidy on its own. Your files never leave the machine. |
qvac-invoice-manager-demo |
Drop folders of invoices and receipts in, including scans and phone photos, and get an accounting table out with the columns you define: those columns become the JSON Schema a local Qwen3 model is grammatically forced to fill, and the arithmetic is rechecked in code rather than trusted to the model. Digital PDFs go through the text path, scans are rasterised and read by Qwen3-VL. Export a CSV for your accounting software. Supplier lists, prices and bank details never leave the device. |
qvac-color-studio |
A Korean-salon-style personal colour analysis, fully on-device. Capture a webcam photo and a local Gemma 4 vision model reads your skin undertone and season, then rates all twelve drape colours in one pass — undertone, season and a comment per colour from a single completion(). Drape any colour under your chin over your own photo and the band recolours instantly, because clicking a swatch is canvas work and never touches the model. Face landmarks come from MediaPipe on the CPU, vendored into the app so nothing loads from a CDN. Your photo reaches the Electron main process and stops there. No cloud, no API keys, offline after a one-time model download. |
qvac-translatepsy-afrislm-demo |
A demo of TranslatePsy-AfriSLM, QVAC's translation model for 19 Sub-Saharan African languages: paste some text or photograph a page, and read it in your language, entirely on-device. The model is paired with a QVAC vision model so a photographed page can be read first, and it goes from one African language to another directly, with no English in the middle. Before downloading anything the app calls assessModelFit to see what the machine can spare, then fetches the size that suits it. The source language is worked out in code rather than asked of a model. Offline once the models are cached, no cloud, no API keys. |
qvac-biomarkers-demo |
Drop in a lab report (a digital PDF, a scan or a phone photo) or a CSV, and read your blood test properly: every marker against its range, a trend per marker, eight category scores you can check by hand, and food ideas and answers from MedPsy 4B, QVAC's medical model, all on-device via @qvac/sdk. Scans and photos are read by the SDK's on-device OCR. The model can only name foods from a knowledge base you can read, the arithmetic is done in code, and a filter stops any dose from reaching the screen. Your results never leave the machine. A demonstration, not a medical device. |
qvac-image-generation-demo |
Type a sentence and get an image on your own computer, in about 11 seconds on an Apple M5 Max. FLUX.2 [klein] 4B runs through @qvac/sdk as one split model (the diffusion model, a Qwen3 4B text encoder and a VAE), in 4 steps at 768 x 768. Six prompt ideas, six styles, a queue for several visitors and a screen that clears itself, for a booth. Offline once the 5.1 GB model is downloaded, no cloud, no API keys. |
qvac-music-desk-demo |
Make music on your own computer with ACE-Step 1.5 through QVAC's audio generation addon. /stand is three picks and a button: a 30 s track in about 11 s, then restyle it, make it longer or take another. The full desk adds a song sheet, a local model that turns a sentence into one, covers, repaint and stems. The caption rules from the model's authors are enforced in code and tested over every combination. Offline once the models are downloaded, no cloud, no API keys. |
qvac-demo-dashboard |
One page to set up, start and stop seven of these demos on one Mac. npm run setup installs each demo and downloads every model it needs with that demo's own @qvac/sdk; after that the dashboard opens each demo with a click, one at a time, and everything runs offline. |
Each app documents its own prerequisites and run steps:
- Open the app's directory (
qvac-{app-name}/). - Read its
README.md. - Follow the install and run instructions there.
QVAC is an open-source, cross-platform ecosystem for building local-first, peer-to-peer AI applications and systems. With QVAC, you can run AI tasks like LLMs, speech, RAG, and more locally across Linux, macOS, Windows, Android, and iOS — or delegate inference to peers using its built-in P2P capabilities.
Learn more at qvac.tether.io, read the docs at docs.qvac.tether.io, or explore the SDK on GitHub.
Licensed under the Apache License, Version 2.0. See LICENSE.
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
Copyright © 2026 Tether Data, S.A. de C.V.