Voice assistant · runs on your GPU
Your Jarvis, 100% local.
A brain, cloned voice, and ears running on your computer. No monthly fees, no API account, and nothing leaves your home.
A brain, cloned voice, and ears running on your computer. No monthly fees, no API account, and nothing leaves your home.
Jarvis Local is an open kit for building a personal assistant you call by voice ("Jarvis, …") that runs entirely on your computer. It answers using your own notes, displayed in a 3D galaxy, speaks with a cloned voice, and records new information when you ask. It’s for anyone who wants an assistant without a monthly fee or sending audio to the cloud. You need a computer with a GPU (an NVIDIA graphics card, a Mac with an Apple chip, or a unified-memory mini PC) and Claude Code or Codex to build it once using the kit’s 8 prompts.
The project explained in a few minutes.
The kit installs three free engines and provides the prompts that have Claude Code (or Codex) build the assistant on top of them. Once it’s ready, it runs without either tool and without internet.
Ollama runs an open AI model (Qwen, Gemma, gpt-oss) on your graphics card and answers using your notes.
VoiceStudio clones your voice from 10–20 seconds of audio. The VoxCPM2 engine speaks Portuguese.
Whisper transcribes what you say right there. Say "Jarvis" and it listens; say "stop" and it goes quiet.
The brain, voice, and ears all stay in memory at the same time, and the brain takes up the most. What matters is how much memory the model can access: the NVIDIA card’s VRAM or the unified memory in a Mac or AI mini PC.
Mac with Apple Silicon (16 GB+, 32 GB ideal)
PC with NVIDIA graphics card (12 GB+ VRAM, 16–32 GB ideal)
Unified-memory mini PC (AMD Ryzen AI Max+, NVIDIA DGX Spark)
VPS / cloud server without a GPU
8 GB laptop without a graphics card
Graphics card with less than 12 GB VRAM (simple chat only)
Download size for each model (4-bit, Ollama library, 11/04/2026). The installer automatically chooses up to gpt-oss:120b; choose the last one in the list with --modelo. Leave ~30% free: the voice and both Whisper engines are estimated to use about 4–6 GB more on the same card.
Also fit: gpt-oss:20b (14 GB, fast), gemma4:26b-a4b (16–19 GB), llama4:scout (67 GB). Tags ending in -cloud run in the Ollama cloud: the installer refuses them. Full table in MODELOS.en.md.
Everything happens on your computer, on local ports. Prompt 07 measures how long each step takes on your machine.
Open-model server with an OpenAI-compatible format. The same code can talk to another endpoint by changing three fields.
Local voice engine. The assistant pins the engine at each startup so no license changes without notice.
The 3D notes galaxy in your browser, bound to 127.0.0.1. You talk through it and see which note the answer came from.
The model name in Ollama, including its size: qwen3.5:9b.
A model compressed to fit in less memory, with little loss in quality.
The address that receives questions, here localhost:11434/v1.
The memory the GPU uses. On Macs and AI mini PCs, it’s shared with the system.
An internet file server. Prompt 08 brings the libraries into the project.
An open library that draws the 3D notes galaxy in your browser.
The installer handles the three engines and checks everything else before starting. If anything is missing, it shows you the command to install it.
NVIDIA graphics card, Mac with an Apple chip, or unified-memory mini PC (see above). On NVIDIA, with the driver installed: nvidia-smi must work.
On Ubuntu/Debian, you also need python3-venv and sudo to install Ollama.
sudo apt install -y python3-venv curl
Or another browser with microphone access.
Only needed once, to build with the prompts.
One evening to install, one weekend to build. Every step ends with proof.
Clone the repository on the computer where the assistant will live.
git clone https://github.com/inematds/jarvis-local.git cd jarvis-local
Shows your GPU, memory, and which model will be selected without installing anything. If there’s no GPU, it stops and explains why.
# Mac or Linux bash instalar/instalar.sh --simular # Windows (PowerShell) powershell -ExecutionPolicy Bypass -File instalar\instalar.ps1 -Simular
Run the same command without simulation: Ollama + the model that fits your memory, Whisper (faster-whisper or mlx-whisper), and VoiceStudio. To choose the model manually: --modelo qwen3.8:27b.
bash instalar/instalar.sh # Mac or Linux bash instalar/instalar.sh --modelo gemma4:12b # force a model powershell -ExecutionPolicy Bypass -File instalar\instalar.ps1 # Windows
On an NVIDIA card, it also installs Whisper’s CUDA libraries (cuBLAS and cuDNN). At the end, it saves the brain tag and the Python interpreter the prompts will use to ~/.jarvis-local/minha-maquina.json.
In VoiceStudio: Model Catalogue → install VoxCPM2 (speaks Portuguese, Apache-2.0 license) and select it. Then go to Profiles → New → Clone with 10–20 seconds of YOUR voice, without music or echo, and name it Jarvis. VoiceStudio is a desktop app; note the version that worked and don’t update in the middle of the project. On Linux, open it with ~/.local/bin/VoiceStudio (it is an AppImage: Ubuntu 24.04 needs sudo apt install -y libfuse2t64, 22.04 needs libfuse2). On first launch it downloads its own engine, which takes a while.
Checks the three engines and measures the brain’s first response on your computer.
python3 instalar/verificar.py # ✓ GPU: … ✓ model downloaded ✓ responded in X s ✓ VoiceStudio running (example)
Open Claude Code (or Codex) in an empty folder and paste prompts 01 → 08 in order, one at a time, waiting for each to finish and prove itself. In prompt 01, fill in the brain tag and Python interpreter from ~/.jarvis-local/minha-maquina.json.
mkdir ~/meu-jarvis && cd ~/meu-jarvis && claude # paste the contents of prompts/01-galaxia-e-cerebro.md
Prompt 07 creates verificar_tudo.py, which measures listening · thinking · speaking on your computer. Prompt 08 turns off Wi-Fi and checks that the list of external connections is empty.
python3 verificar_tudo.py # listen 0.6 s · think 1.9 s · speak 0.8 s · total 3.3 s (example) # 14 passed, 0 failed, 0 warnings
"Jarvis, what did I note about client X?" · "Jarvis, note that the meeting moved to Thursday." · "Switch to the large model." · "Keep everything local."
Each prompt is a complete request: what to build, the rules it must follow, and how to prove it works. Change only what’s in square brackets.
Two things the kit delivers: an answer flying to the note it came from, and an assistant working with the internet turned off.
What the kit teaches you to do right, and where to go next.