Vui
fluxions-ai/vuiReviewed by AI
Checked by RepoPicks Radar on
What is Vui?
Two things in one repository: a small speech model that reads text out loud and can copy a voice from a sample, and a complete voice assistant built on it. You open a page on your own machine, talk into the microphone, and it transcribes, asks a language model you run (Ollama or vLLM) and answers out loud, with interruption, memories and timers. It is built for Linux with an NVIDIA card and installs through Docker or a shell script; the Apple Silicon side runs the speech demo but the full assistant there is unfinished. There is also a C version that speaks on a plain processor.
Vui is free and open source (Apache-2.0) and can replace ElevenLabs and ChatGPT.
How easy is it to install Vui?
D · terminal, several tools or a specific graphics card · macOS
- macOS D
- Linux D
Linux with an NVIDIA card is the intended setup and still a D: you need Docker with the Compose plugin, the NVIDIA Container Toolkit (installed with sudo), Ollama running on the machine with a model pulled, and the ffmpeg shared libraries, or you run the shell installer and let it do it. On a Mac it is also a D and less complete: the README says the speech demo works on Apple Silicon but "the rest of the MLX stack is WIP", including the streaming server. Windows is not mentioned.
Coming soon: install in 1 click with the RepoPicks app.
Vui on GitHub
- Stars
- 776on Nov 10, 2026
- Downloads
- —all releases
- Growth
- —no gain in the last reading
Stars over time
Facts
- Runs on
- Linux, macOS
- You need to install
- Python, Docker, NVIDIA graphics card, FFmpeg
- To build from source
- Python, Docker, NVIDIA graphics card, FFmpeg
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- Anthropic (or a Claude Code subscription), only for the optional task helper that reads mail and calendar, Serper, Brave or Tavily, only for spoken web searches
- Cost
- Free · Apache 2.0 with the weights included and commercial use allowed. The authors also sell a hosted version: the README calls this repository "the open core" and says their production API "ships ongoing model updates and a more advanced turn-taking system".
- License
- Apache-2.0 (Permissive)
- Can replace
- ElevenLabs, ChatGPT Plus
- Last change on GitHub
- Oct 10, 2026
Vui risk signals (7)
- AttentionInstall command runs a script from the internet
Install command pipes a remote script into a shell: curl -LsSf https://astral.sh/uv/install.sh | sh
automatic check
- Attentionvoice cloning of any speaker
The README says "The model can clone arbitrary voices — upload a sample in the demo UI (or drop a .wav into prompts/) and it will follow that speaker", and from the command line "Point --prompt at any .wav. No transcript needed". The project's own note on misuse says the checkpoint "is fine-tuned to a curated voice set in part to make these misuses harder, but it is not a substitute for your own judgment". Copying someone's voice without permission can break the law.
AI review
- Attentioninstall one liner runs remote code
The documented install is "curl -fsSL https://install.fluxions.ai | bash", which the README says "Clones into ~/vui, auto-detects Docker vs. native, installs deps (uv, ffmpeg libs, Claude Code CLI), and launches the stack". Whatever that address serves at the moment you run it is executed on your machine without review.
AI review
- Attentionoptional helper gets your mail and claude login
The compose file starts a "claude-task" sidecar by default and mounts your login into it: "${HOME}/.claude:/root/.claude:ro" with the comment "Claude Code subscription credentials (after claude login on host)", or an ANTHROPIC_API_KEY. The README says it handles "Gmail / Calendar / Drive / Slack reads, web research" through your existing Claude Code MCPs, so a spoken sentence can read your mail.
AI review
- Attentionserver has no login and tunnels are documented
The services run with "network_mode: host" on ports 8080, 8443 and 8642 with no password mentioned anywhere, and the phone section suggests "cloudflared tunnel --url http://localhost:8080 — one command, HTTPS, no account". That publishes a microphone-driven assistant, your memories and the task tools to anyone with the address.
AI review
- Infowhat you say is written to disk
The assistant keeps facts about you between sessions: "Memories are loaded from ~/.vui/memories.json on startup and rewritten on every add/remove", and the compose file mounts "./debug_dump:/app/debug_dump" with the comment "Optional: write conversation logs to host for inspection".
AI review
- Infoweights download separately and one set is unpublished
The README says "The TTS weights are not pulled here — they download from Hugging Face on first render", and the compose file overrides the command with the comment "load a local checkpoint instead of fetching vui-nano.safetensors from huggingface.co/fluxions/vui (not yet published)" and "Drop this line once the HF repo has the v2 weights".
AI review
Evidence is shown in its original language.
What Radar could not check about Vui
- Whether the install finishes on your machine, as it builds Python, CUDA and ffmpeg pieces at install time.
- The speed and accuracy tables (9x real time on a 4090, word error rates), measured by the authors.
- How natural the voice and the turn-taking feel in a real conversation, and how well a cloned voice holds up.
- Whether the checkpoints the code asks for are actually published on Hugging Face right now.
- Real disk, memory and video-memory use once the speech model, the speech-recognition model and a language model are all loaded.
- What the hosted demo at the authors' site does with the audio you send it.