Laya Serve for macOS: a menu bar app with an OpenAI-compatible endpoint for n8n

# Laya Serve for macOS

A macOS menu bar application that serves [Laya](https://github.com/NandhaKishorM/laya)
over HTTP with an OpenAI-compatible interface. Point the n8n **Text Classifier** node at
it and classify emails on your own Mac.

Laya is a non-autoregressive decision engine from Convai Innovations. It answers typed
questions in one forward pass of about 40 ms. It does not generate text, so it cannot
hallucinate an answer that is outside your category list.

The weights ship inside the application. It needs no network, no Hugging Face account
and no download at run time.

## What it does

- Runs a local HTTP server that speaks the OpenAI chat completions protocol.
- Translates each n8n classification request into Laya typed questions.
- Returns the exact JSON object that the n8n node expects.
- Holds the model weights inside the application, so it works with no network.
- Unloads the model after an idle period, and loads it again on the next request.
- Lives in the menu bar. It has no Dock icon and no window.

## Requirements

- Apple Silicon. PyTorch has no current macOS x86_64 wheel.
- macOS 13 or later.
- About 1.5 GB of disk. That holds CPython, PyTorch and the weights.
- About 1.3 GB of memory while the model is loaded, and about 60 MB when it is not.

## Install

With Homebrew:

```bash
brew install --cask chrisns/laya-mac-serve/laya-serve
```

The application is ad-hoc signed, not notarised, so macOS blocks the first launch.
Remove the quarantine flag:

```bash
xattr -dr com.apple.quarantine /Applications/LayaServe.app
```

By hand:

1. Download the disk image from the
   [latest release](https://github.com/chrisns/laya-mac-serve/releases/latest).
2. Drag `LayaServe.app` to Applications.
3. The application is signed ad-hoc, not notarised. Remove the qua