laya-mcp — on-device typed decisions in ~10ms

Laya-MLX as an MCP server (and plain-HTTP API). It answers typed decisions — choose an option / score a rubric / is this true? — on your own machine (Apple Silicon / MLX), with no cloud and ~10ms after warm-up.

Not a chatbot. No token-by-token text, no JSON that can break. One forward pass returns a structured answer you can branch on. Ideal for routing, triage, classification, lead scoring, and guardrails where an LLM would be slow, expensive, or would ship your data to the cloud.

⚡ Runs on Apple Silicon (M-series) out of the box. Multi-language checkpoint (EN/DE + more) by default, so it triages real inbox/ticket text.

Tools

Related MCP server: apple-fm-mcp

Quick start (MCP)

One-liner via uvx (no install):

uvx --from laya-mcp laya-mcp

Claude Desktop / Cursor / any MCP client → add:

{
  "mcpServers": {
    "laya-mlx": {
      "command": "uvx",
      "args": ["--from", "laya-mcp", "laya-mcp"]
    }
  }
}

First call downloads the model (~30s, cached after). After that: ~10ms.

Raw stdio (no MCP client)

Newline-delimited JSON-RPC 2.0:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"laya_decide","arguments":{
     "state":"Der Kunde wurde doppelt berechnet, bitte erstatten.","question":"Which team handles this?","options":["billing","technical","sales"]}}}' \
  | uvx --from laya-mcp laya-mcp
# -> id 2: {"choice":"billing","confidence":0.9999,"probabilities":{"billing":1.0,...},"ms":...}

HTTP API (n8n / Make / curl)

For nodes that just POST JSON — no MCP client required:

uvx --from laya-mcp laya-http          # listens on 127.0.0.1:8901

curl -s localhost:8901/decide -H 'Content-Type: application/json' -d '{
  "state":"Wir suchen einen Partner für die Digitalisierung unserer Buchhaltung.",
  "question":"Pursue this lead?","options":["pursue","maybe","drop"]}'
# -> {"choice":"pursue","confidence":0.74,"probabilities":{"pursue":0.92,...},"ms":1.8}

Environment

Why on-device

• Privacy / GDPR — data never leaves the machine.

• Latency — ~10ms = real-time routing, not "fast".

• Cost — ~zero marginal; millions of decisions/day are practical.

Works with

Tested end-to-end against a Go MCP gateway (stdio transport, tool policy, per-key budget + kill switch, metering) — so it slots into an existing agent/LLM gateway as one more, very cheap, local tool.

License

MIT