01
The app
Listens & speaks
Native on macOS, iOS and Android. Microphone, echo cancellation, end-of-turn detection, speech recognition, text to speech, barge-in. Your voice never leaves the device.
Self-hosted · Voice-first · Open
Your own
assistant.
Not an account on somebody’s service. Not a wrapper around an API key someone else holds. MyAssistant runs on a machine you own, remembers in files you can read, and answers to you.
You talk to it. It talks back — out loud, while it works.
Drop a <video> or embed here
- macOS
- iOS
- Android
- Self-hosted
- Yours, end to end
“The idea is old and the joke is obvious: everybody wants the thing Tony Stark has. The point of this is that you do not need to be him. A phone, a machine you already own, and an evening.”
Architecture
Three parts. No one of them owns you.
The app listens and speaks. The harness thinks. The vault remembers. Deliberately separate, so you can replace any piece without losing the others.
02
The harness
Thinks
A warm agent session behind an HTTP and SSE socket, running wherever you want the assistant to live: your Mac, or a server you reach over SSH. It streams sentences as they are written, and every tool call that needs permission comes back to you first.
03
The vault
Remembers
Markdown with frontmatter and wikilinks — one file per note, in
numbered folders. Obsidian opens it. So does any text editor, so does
grep. The assistant maintains it, and it outlives any version
of this software.
That last part is the one people miss. The memory is not rows in a database belonging to a company. It is a folder. You can read it, edit it, back it up, move it to another machine, or throw it away — and the assistant carries on from whatever it finds there.
Voice, done properly
It starts talking before it finishes thinking.
- Streams sentences as they are written — the voice begins while the thought is still forming.
- Barge-in — interrupt it mid-sentence. It goes quiet and listens.
- End-of-turn detection — no robot timeouts, no “I’m still listening” theater.
- Echo cancellation and on-device speech recognition — your voice never leaves the device.
Replace with a real capture
Replace with a real capture
Projects
Work you don’t wait for.
Register a repository and the assistant gets a persistent session in it. Ask it something and it answers; give it a job and it takes a git worktree of its own, works on a branch, and tells you when it is done — by speaking up, not by waiting for you to ask.
Run several at once. They cannot collide, because each has its own checkout. The assistant is the chief of staff, not the engineer: it relays, it summarises, and it never claims a job is finished that it has not checked.
Bring your own brain
A good voice in front of any rig you already run.
The harness contract is an HTTP and SSE socket with a documented event vocabulary. Claude Code today, in full — files, commands, git repositories, and it asks before each one. The MyaiOne Rig brings the same assistant to sixty-odd providers. If anything you run can stream sentences and ask for permission, it can wear this voice.
Claude Code
MyaiOne Rig
Hermes
OpenClaw
Any streaming harness
Privacy, as an architecture decision
There is no cloud account for it to leak.
Your voice stays put
Speech recognition happens on the device. Nothing is shipped to us, because there is no “us” service in the loop.
Your memory is a folder
Plain Markdown you can read, version, and back up. Delete the app tomorrow and your notes are still notes.
Your machine decides
The harness runs where you choose. Tool calls that matter come back to you for approval — it asks before it acts.
Getting started
A machine you already own, and an evening.
1
Point the app at a machine
Your Mac, or a server you reach over SSH. It puts the harness there, creates the assistant’s folder, scaffolds its vault, starts it, and connects.
2
It interviews you
There is no setup wizard. The assistant interviews you — and its own instruction file is the interview. When it finishes, it writes its own instructions and is never a wizard again.
3
Start talking
Ask it things. Give it jobs. It answers out loud, works in the background, and speaks up when something is done.
One honest note: the harness needs its engine installed and logged in on that machine. That is the one thing the app cannot do for you — and it says so plainly rather than failing halfway.
Up close
What it looks like.
screenshot
screenshot
screenshot
FAQ
Fair questions.
Do I need an account with you?
No. MyAssistant is self-hosted. It runs on hardware you own, and there is no service of ours sitting between you and your assistant.
Where does my voice go?
Nowhere. Speech recognition happens on the device. The transcript goes to the harness you chose to run, on the machine you chose to run it on.
What are its memories made of?
Markdown files in folders — frontmatter, wikilinks, plain text. Obsidian opens them; so does any editor. Back them up like any other files, because that is what they are.
Does it just chat, or can it actually do things?
It drives real engines: it reads and writes files, runs commands, and works in git repositories — and every tool call that needs permission comes back to you first.
Which platforms are supported?
Native apps for macOS, iOS and Android, built on one shared core. The harness runs on macOS or on any Linux server you can reach over SSH.
Is it free?
MyAssistant is free, open-source software. Your only costs are the hardware you already own and whatever models or services you yourself choose to connect.