Skip to main content
Glama
README.md
# codex-laya-mcp

Local, read-only Laya typed decisions for Codex over Model Context Protocol.

This project is intentionally not a replacement model for Codex. It exposes bounded classification,
scoring, truth-probability, batch decision, and token-budget planning tools backed by `laya-mlx` on
Apple Silicon.

## Included tools

- `laya_plan`: exact prompt-budget preflight without inference.
- `laya_decide`: batched choice, score, and truth-probability decisions.
- `laya_classify`, `laya_score`, and `laya_check`: focused convenience tools.

Every inference result includes the full distribution, uncertainty metrics, model metadata, and an
advisory answer/review/abstain policy. Inputs that Laya would silently truncate are rejected.

## Development

```bash
uv sync --extra dev
uv run ruff check .
uv run pytest
uv run codex-laya-mcp --help
```

The first real model call downloads the configured checkpoint unless it is already in the Hugging
Face cache. Unit and MCP contract tests use a deterministic fake engine and do not download weights.

See [CODEX_MCP_PROJECT_DESIGN.md](CODEX_MCP_PROJECT_DESIGN.md) for the complete design and
[docs/OPERATIONS.md](docs/OPERATIONS.md) for installation, Codex connection, and rollback.