laya-mcp: a local MCP server exposing laya-mlx to any MCP client

# laya-mcp

`laya-mcp` runs the [Laya](https://github.com/mizorewww/laya-mlx) decision model locally and exposes it as an [MCP](https://modelcontextprotocol.io/) server. Use it to classify text, score text against a rubric, or answer yes/no questions without sending inference input to a remote service.

Laya is a decision model, not a chat model. It returns probabilities for structured questions such as “Is this urgent?” or “Which category fits this message?”. Inference uses the Apple Silicon GPU through MLX.

## Requirements

- macOS on Apple Silicon
- Python 3.12 or newer
- [`uv`](https://docs.astral.sh/uv/)

## Install

```sh
git clone https://github.com/wsargent/laya-mcp.git
cd laya-mcp
uv sync
```

The model checkpoint downloads from Hugging Face on the first inference request and is cached locally. The server starts without loading the model; the first request may take longer than later requests.

## Use as an MCP server

Start the stdio server from the project directory:

```sh
uv run laya-mcp
```

Register the server in an MCP client with the command below. Use the absolute path to the Python executable in this project's `.venv`:

```json
{
  "command": "/absolute/path/to/laya-mcp/.venv/bin/python",
  "args": ["-m", "laya_mcp.server"]
}
```

See [client setup](docs/client-setup.md) for Claude Desktop, Polytoken, and the shared daemon.

## Tools

The server provides five tools:

| Tool | Purpose |
| --- | --- |
| `laya_decide` | Answer caller-defined `choice`, `score`, and yes/no (`noul`) questions about a string, object, or list. |
| `laya_triage` | Classify a support message by intent, urgency, frustration, refund request, and churn risk. |
| `laya_guard` | Check a prompt for jailbreak, prompt-injection, sensitive-data, harm, and topic signals. |
| `laya_modera