Nanojev MCP — DevOps skill for Claude Code

NanoJev MCP

[简体中文](README_CN.md)|[English](README.md)

Wraps [NanoJev](https://github.com/TianyuCodings/NanoJev)'s decision service as an MCP server, letting any MCP-capable agent (Claude Code, ccwitch, Cline, Cursor, ...) query the NanoJev model in natural language.

agent ──stdio(MCP)──> mcp_nanojev.py ──HTTP──> NanoJev repo's serve_decisions.py :8765 (model loaded once)

The MCP process is a thin proxy that never loads the model; multiple agents can share the same inference service simultaneously.

Prerequisite: Install NanoJev

git clone https://github.com/TianyuCodings/NanoJev.git
cd NanoJev
python -m pip install -r requirements-toy.txt huggingface_hub

Download the checkpoint (~2.3 GB; see the NanoJev README for details):

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="C-Tianyu/NanoJev",
    revision="unified-games-v1",
    local_dir="checkpoints/NanoJev-unified",
    allow_patterns=["best.safetensors", "config.json", "tokenizer/*", "backbone_config/*"],
)

Setup

**1. Download this package** (clone / unzip to any directory, e.g. `D:\nanojev_mcp`)

**2. One-command configuration**

python setup_mcp.py

The script will: install the `mcp` SDK -> probe the inference service -> register `nanojev` into Claude Code (user scope) -> print JSON for other clients. Common options:

python setup_mcp.py --service-url http://127.0.0.1:9000/api/evaluate   # non-default port
python setup_mcp.py --python D:\NanoJev\.venv\Scripts\python.exe       # use NanoJev's venv
python setup_mcp.py --skip-claude                                      # only print JSON, leave claude config untouched

**3. Start the NanoJev inference service** (in the NanoJev repo directory)

python scripts/serve_decisions.py --checkpoint-dir checkpoints/NanoJev-unified --web-root web --port 8765 --disable-native-triton

Done. Just ask your agent in natural language, e.g.:

Use nanojev to decide: the