# laya-mcp

[Laya](https://github.com/NandhaKishorM/laya) as an **MCP server**: typed
decisions (`choice` / `score` / `noul`) with real probabilities in ~20-40 ms,
on your own machine. No LLM calls, no tokens, no API keys, 100+ languages.

Same tool interface as `jev-local`, so it is a drop-in replacement wherever
those tools are used.

## Install

```bash
pip install git+https://github.com/jerepaira/laya-mcp.git
```

This pulls `laya` (which pulls `torch`) — for NVIDIA GPU you get CUDA
support out of the box; on CPU-only machines it just runs slower
(~15 s for the first call, milliseconds after warmup).

Model weights (~2 GB) download automatically from
[HuggingFace](https://huggingface.co/convaiinnovations/laya) on first use.

## Use with your agent

**opencode** (`~/.config/opencode/opencode.json`):

```json
"mcp": {
  "laya": {
    "type": "local",
    "command": ["laya-mcp"],
    "enabled": true
  }
}
```

**Claude Code** (`claude mcp add`):

```bash
claude mcp add laya -- laya-mcp
```

Any other MCP client: stdio command `laya-mcp`.

## Tools

| tool | what it does |
|---|---|
| `decide(state, questions)` | answer any set of typed questions at once |
| `classify(text, labels, instructions?)` | pick one label + confidence + scores |
| `score(text, criteria, instructions?)` | rate on an ordered scale + distribution |
| `check(state, question)` | yes/no → `P(yes)` in `[0,1]` |

Example — route a support message:

```json
classify("me cobraron dos veces, quiero la devolución",
         ["billing", "technical", "sales", "other"])
→ {"label": "billing", "confidence": 0.845, "scores": {...}}
```

## Config

| env | default | meaning |
|---|---|---|
| `LAYA_DEVICE` | auto (`cuda` if available, else `cpu`) | force `"cpu"` if VRAM is tight |
| `LAYA_MODEL` | auto-router per request | pin `"english"`, `"multilingual"` or `"typed-decisions"` |

Tip: Spanish text routes to the multilingual checkpoint. Pinning one model
with `LAYA_MODEL` avoids checkpoint reloads when requests mix languages.

## Accuracy notes

Out of the box it is fast and decent, not magic — wording of the question
matters a lot and thresholds should be calibrated on your own labeled
examples before trusting the probabilities. Measure, don't assume.

## Credits

Decision model by [Nandakishor M (Convai Innovations)](https://github.com/NandhaKishorM/laya),
Apache-2.0. This repo is only the MCP wrapper (MIT).