A voice assistant you talk to on your own Mac or Windows PC. A shader orb, a 3D graph of your files, real tools and Claude Code agents on command, all on your Claude subscription with no API key. Built by Andy Hafell.
What you build
| Part | What it is | What it does |
|---|---|---|
| Eyes | A shader orb in an app window | Shows idle, thinking and speaking, driven by the engine |
| Map | A 3D force graph of your files | Lights up with whatever he just searched for |
| Brain | Claude Code CLI on your subscription | Answers as a dry British butler, with a real shell |
| Voice | faster-whisper or whisper.cpp, Kokoro voice | Speech in, speech out, all local |
| Hands | Tools for files, web, screen, camera, Telegram, email | Each refuses out loud when it is missing what it needs |
| Agents | Claude or Codex agents in tmux or Windows Terminal | Started, briefed and read back by voice, four at most |
Where this comes from
AF Jarvis was built by Andy Hafell, and he gave permission to share it here. The engine, the orb, the 3D graph, the tools, the agent launcher and the installer are all his work. Nothing in this pack was rebuilt or reworded by us beyond a credits file. If you want more apps like this, and to see how Andy builds them, join his community: skool.com/aimate.
What you build
You install a finished app, you do not build one from scratch. You press Talk, ask him to pull something up, and the graph on the right filters down to the files he actually found. He answers in one or two dry sentences, in a British voice, and he calls you sir.
He asks before deleting anything or sending anything. He can start Claude or Codex agents for you, four at most, and he deliberately cannot kill one: a misheard sentence that closes a window would cost the work inside it.
If you would rather understand the build than just run it, PROMPT_PACK.md in the zip is Andy's full prompt sequence: one baseline prompt that builds a working first draft from an empty folder, then sixteen more, one per feature, plus every trap that cost him a day.
Do you need this? Be honest first
This is for you if you already pay for Claude Pro or Max, you like talking to your machine, and you want to see how a voice assistant is wired end to end: speech in, a model with real tools, speech out, and a UI that shows what it did.
Close the tab if you want a phone assistant, a cloud service, or something that runs without Claude Code. There is no API-key mode, on purpose: a voice assistant on a metered key spends money while you talk to it.
What you need
- The machine. A Mac on macOS 13 or later, or a Windows 10 or 11 PC. Screen reading on the Mac needs macOS 14.
- Python 3.10 to 3.13. Not 3.14 yet: the Kokoro voice has no build for it. The Windows installer finds a supported Python on its own if you have several.
- Claude Code, logged in with a subscription. On Windows use the native build, not the npm one.
- On Windows, Git for Windows and Windows Terminal.
- The time. About 15 minutes, most of it the one-time downloads: the Kokoro voice (about 340MB) and the speech model (about 140MB).
What is in the zip
AF-Jarvis/
README.md what it is, how it works, the traps
INSTALL.md every dependency, what dies without it, Mac and Windows
FEATURES.md what is done, what is partial, and why
PROMPT_PACK.md Andy's prompts to build it yourself from an empty folder
CREDITS.md Andy Hafell, and the open source pieces inside
install.sh the Mac installer (install.py on Windows)
server.py the engine: voice, ears, brains, tools
public/orb.html the orb and the 3D graph
app/ the native Mac app (Swift)
wake_models/ the "Hey Jarvis" wake word, runs offline
On a Mac run ./install.sh. On Windows, in PowerShell, run py install.py. Both print a report of what your machine can do, and anything marked with an x is explained in INSTALL.md with the fix.
What was verified and what was not
Verified on a Windows 11 PC on 2026-09-23 with Python 3.12.10: py install.py built its own virtual environment, installed every package, downloaded the voice and speech models, and reported brain, ears, voice, wake word, agents and graph all available. The app window opened on port 7915. A typed question came back in character in 2.6 seconds, and the Kokoro voice spoke out loud.
Not verified by us: the Mac install (Andy built and measured it on an Apple Silicon Mac), the wake word with a real human voice, and the Telegram, email, invoice and camera tools. Each of those needs something that is yours, like a bot token, a Gmail token or an invoice template, and each refuses out loud and names what it wants until you supply it.
Read the prompt before you run it
The installer is a script that runs on your machine, and the assistant has a real shell. Hand install.py or install.sh to Claude first and ask what it will install and where. Do the same with any prompt in PROMPT_PACK.md before you build with it. A question costs a minute. A surprise costs an afternoon.
