There is a lot of very good open work in this space right now, and almost none of it is something you can talk to. NVIDIA put out a speech model that runs on your laptop. An independent researcher put out an 82-million-parameter voice. Nous Research put out an agent runtime under MIT. Every one of those is excellent, and every one of them ends at a terminal or a Python import.
Memoe is those pieces, assembled, behind a face you can talk to. No chat box: it lives in your notch, in every app on your Mac. You say "Hey Moe", it answers out loud, and it does the work. https://dub.sh/nxzivCr
Here is what is in it and where each part comes from.
Speech, both directions, on the Neural Engine
Speech to text is NVIDIA's Parakeet (parakeet-tdt-0.6b-v3), converted to CoreML and running on the Apple Neural Engine. End-of-utterance detection is a second, smaller Parakeet streaming model, so Moe knows you have finished a sentence without waiting on a timer. Speaker diarization is in there too, which is what makes meeting notes name the right person.
The voice that answers is Kokoro, 82 million parameters, also CoreML on the ANE.
Nothing here is a cloud call. That is why it is free, why it is fast enough to interrupt, and also why Memoe is Mac-only. This does not work as an Electron app.
The harness is Hermes
The agent runtime is Hermes Agent from Nous Research, MIT licensed: model routing, tools, memory, the plugin surface. It runs as its own process that Moe launches and talks to over HTTP on 127.0.0.1, in a URLSession with the proxy dictionary explicitly emptied, so a system-wide proxy never sees your mail or your memory.
Hermes is genuinely good and almost nobody has met it, because meeting it means a terminal.
The model is the one you already pay for
Memoe brings no model of its own. It drives the official Claude Code or Codex CLI as a long-lived child process, or an API key, or a local model in LM Studio or Ollama.
So if you already pay for Claude or ChatGPT, Memoe is free and you get exactly the model you already trust. Same quality, out of the chat window. One subscription, not two.
What I actually built
The part where those stop feeling like pieces.
A window that is not a window: the island under the notch, which grows out of the notch when Moe is thinking and shrinks back when it is done. Wake word and push-to-talk that work from any app without stealing your cursor, so you can keep typing somewhere else while it works. Turn-taking that does not cut you off at a breath. A memory that is one plain text file on your disk, where every line says where it came from, so you can open it, fix a line, or delete a line. A ladder for driving the Mac that goes structured data, scripting, menus, keyboard shortcuts, accessibility elements, and only then pixels. Connections to mail, calendar and messages that are already set up when you first speak to it.
And two rules that came before any feature: before Moe sends anything as you it shows you what is going out and waits for a yes, and if you touch the trackpad while it is working it stops.
Try it
macOS 15 or later, Apple silicon. Signed, notarized, updates itself. Free.
https://dub.sh/nxzivCr or brew install --cask izzy-gottz/moe/moe
It is beta. If it breaks, tell Moe and it reaches me.
Top comments (0)