Laya Setup Guide: A Free Local Decision Engine for Claude Code

Most AI models write text. Laya does not. It is a "System 1" decision engine: you hand it some text (an email, a ticket, a document) and a set of typed questions, and it returns the answers with a calibrated confidence for each, in a single forward pass. Because it never generates text, there is nothing to parse and nothing to make up.

It has three question types:

• choice: pick one label, like which department should handle this ticket.
• score: rate on a scale, like how urgent is this: not urgent, soon or blocking.
• noul: yes or no, returned as the probability that the answer is yes, like does this customer threaten to cancel.

It works in 100+ languages, and a built-in router picks the right model for each request. The README measures 33 ms for one question and 7.2 ms per question when batched, on a T4 GPU.

• Claude Code installed.
• Python 3.10 or newer.
• A normal computer works: Laya runs on CPU, Apple Silicon (MPS) or an NVIDIA GPU. A GPU is just faster.
• Internet the first time you run a prediction, so it can download the model from Hugging Face.

Easiest: paste this into Claude Code:

Or do it yourself:

Then try the built-in command line. The first line only routes and works offline. The second gives full answers from a ready-made triage question set (it downloads the model the first time):

Laya ships an optional MCP server, so Claude Code can call it as a tool for typed decisions. Install the extra:

Then add it to Claude Code. The LAYA_DEVICE=cpu setting is the README's own example; drop it to let Laya pick your GPU automatically:

Restart Claude Code and ask it to use Laya:

• Support triage: route tickets by department, urgency and churn risk.
• Inbox sorting: lead, spam or needs a human, in any of 100+ languages.
• Prompt guardrails: flag jailbreak or injection attempts before they reach your agent.
• Model routing: decide whether a request needs a big model or a small one, so you spend less on easy tasks.

Out of the box Laya works zero-shot, but the README is clear that fine-tuning on your own domain is where accuracy jumps. On their typed-decisions benchmark of 2,000 decisions, the fine-tuned model scored 0.766 accuracy against 0.362 for the base English model. The repo includes a Kaggle notebook and an Apple Silicon script that run the whole loop.

• 33 ms is on a GPU. That number was measured on a T4 GPU. On a normal CPU it is slower: the README's own table shows a few hundred milliseconds per request.
• Zero-shot accuracy can be low on your domain. Until you fine-tune on your own decisions, check its answers carefully.
• Long inputs get less reliable. In the README's test, results stayed strong up to about 4,000 tokens of text and varied beyond that.
• Test on your own data before automating anything that matters. The confidence scores help, but pick your thresholds from results on your real data.

1. 1pip install laya and run the triage preset to see it work.
2. 2Add the MCP server to Claude Code with claude mcp add.
3. 3Ask Claude Code to use Laya on a small batch of your real emails or tickets.
4. 4Check the answers by hand before you trust it with anything automatic.