laya-mcp

MCP server for Laya typed decisions — noul (yes/no), choice, score — wrapped so it survives contact with a server.

npx -y laya-mcp --help

This npm package is a launcher, not an implementation. Laya is a PyTorch model, so the server itself is Python. Installing this package gets you the Node entry point and nothing else; you also need the Python side:

pip install "laya-mcp[mcp]"

npx laya-mcp finds that interpreter, hands stdio to it, and passes the MCP stream through untouched. It does not install anything for you — fetching torch (~2 GB) and a checkpoint (~650 MB) as a side effect of running a command is not something a launcher should do.

If it cannot find your Python

On Windows, python on PATH is often the Microsoft Store execution alias, which is a stub: it cannot import anything, and it sits ahead of your real interpreter. The launcher therefore reports what it tried rather than claiming the package is missing. Point it at the right interpreter either way:

# name it outright
set LAYA_MCP_PYTHON=C:\path\to\python.exe        # Windows
export LAYA_MCP_PYTHON=/path/to/python           # macOS, Linux

# or just activate the virtualenv you installed into — VIRTUAL_ENV is honoured

LAYACORE_PYTHON is still read. 0.1.0 and 0.1.1 printed that name in the error message, and a launcher that stopped honouring the variable its own previous version told you to set would be worse than one carrying the alias.

Status: 0.1.1, work in progress. The Python core is implemented and covered by CI; the end-to-end path through live agent harnesses is not yet covered. Interfaces may move before 1.0.

Why it exists

Laya truncates things silently and reports none of it:

• It cuts the state from the end. A long document loses its tail — for a contract or an email thread, often where the answer was — and the model answers about the surviving prefix at full confidence.

• It shortens options until labels are indistinguishable. Options share a fixed per-question token budget; past a point every label gets ~4 tokens. This is the documented cause of its collapse on high-cardinality choices.

• Its confidence is not accuracy. It is a normalised entropy, low whenever probability is spread out even when the top option is right, and high on a confident wrong answer.

• It demotes itself to CPU on a CUDA OOM, permanently, with no flag set — roughly 10-15x slower, and nothing in the response says so.

This package adds a preflight that reports what would be cut, structured errors that name the offending question, an honest confidence contract, and a health surface that admits a demotion.

Related MCP server: jev-mcp

Usage

laya-mcp serve      # warm HTTP sidecar on 127.0.0.1:8787 (loads the model once)
laya-mcp mcp        # MCP over stdio
laya-mcp doctor     # what is installed, and whether the GPU really works
laya-mcp install    # register with whichever agent harness you have

install handles the fact that there is no portable way to register an MCP server: Claude Code (mcpServers), Codex ([mcp_servers.<name>]), opencode (mcp, with an array command), OpenClaw (mcp.servers) and Hermes (mcp_servers) each get the right shape written to the right file, merged and backed up rather than overwritten. pi has no native MCP support and is reported as unsupported.

Prefer serve plus --sidecar over hosting the model in each stdio process: a harness spawns one server per session, and loading a 650 MB checkpoint per session is the main reason a Python-backed MCP server feels slow.

Honest limits

Upstream's own numbers, repeated because an integration that implies otherwise is lying to you: the base checkpoints are near chance zero-shot on typed decisions (0.362 against a 0.461 majority-class baseline); score is the weakest primitive (35% vs 70% in independent measurement); raw calibration error is 0.466 before temperature fitting; and one fixture run answered "A" on 46 of 50 multiple-choice items.

Calibration makes a probability honest; it cannot make a model right.

Licence

Apache-2.0. Laya is Apache-2.0 by Convai Innovations. This is an independent integration, not affiliated with or endorsed by that project.