A tiny offline coding assistant that suggests refactors or explains code using a local Gemma‑4 model

Quick Start • Features • Examples • Contributing

What is this?

Agentic Coding MCP Server is a command‑line tool that runs completely offline. It loads a locally stored Gemma‑4 GGUF model and, given a short code snippet, either returns a refactored version or a plain‑language explanation. It is built in Python 3.10+, uses torch and transformers for inference, and sentencepiece for tokenization. Ideal for solo developers who want fast, private AI help without any internet round‑trip.

Usage example – refactor:

$ gemma_code_helper.py --mode refactor --file example.py
def foo():
    # unused variable
    x = 1
    return x

Output:

def foo():
    return 1

Usage example – explain:

$ gemma_code_helper.py --mode explain --file example.py

Output:

The function foo returns the value of variable x after assigning it 1. It contains no side effects.

Features

Feature	Description
Offline inference	Runs entirely on local CPU/GPU, no network calls
Refactor mode	Suggests code improvements (dead code removal, loop simplification, PEP‑8 compliance)
Explain mode	Provides concise plain‑language description of snippet behavior
Multiple model formats	Accepts any GGUF file; defaults to gemma-4-q4_k_m.gguf
Flexible prompting	Control instruction prefix via --mode flag
Minimal dependencies	Only torch, transformers, sentencepiece; no uv or RPC server
Zed editor agnostic	Works with any editor; output is plain text
MIT license	Permissive open‑source license

Quick Start

1. Clone the repository: git clone https://github.com/yourusername/agentic-coding-mcp-server.git
2. Create a virtual environment: python -m venv venv && source venv/bin/activate
3. Install dependencies: pip install -r requirements.txt
4. Place a GGUF model file (e.g., gemma-4-q4_k_m.gguf) in the models/ directory.
5. Run the helper: python gemma_code_helper.py --mode refactor --file my_script.py
6. View the refactored snippet printed to stdout

Examples

Title: Delete unused variable
Command: python gemma_code_helper.py --mode refactor --file dead.py
Input (dead.py):

def process():
    unused = 5
    result = 10
    return result

Output:

def process():
    return 10

Title: Describe what a function does
Command: python gemma_code_helper.py --mode explain --file greet.py
Input (greet.py):

def greet(name):
    """Return greeting."""
    return f"Hello, {name}"

Output:

The function greet takes a string name and returns a greeting message that starts with Hello, followed by the provided name.

File Structure

Agentic Coding MCP Server: Local Gemma 4 powered coding assistant integrated with Zed editor/ ├── gemma_code_helper/ │ ├── init.py │ ├── main.py │ ├── cli.py │ ├── engine.py │ ├── formatter.py │ └── prompt.py ├── tests/ │ ├── init.py │ ├── test_cli.py │ ├── test_engine.py │ ├── test_formatter.py │ └── test_prompt.py ├── models/ # Place your GGUF files here ├── assets/ # Infographic and other media ├── requirements.txt ├── requirements-dev.txt └── README.md

Tech Stack

Technology	Purpose
Python 3.10+	Core language
torch ≥2.3	Model inference and tensor ops
transformers	Tokenizer/model loading for GGUF
sentencepiece	Subword tokenization
argparse	CLI argument parsing
pathlib	Filesystem handling
MIT License	Licensing

Contributing

• Fork the repo and create a branch.
• Make changes with tests.
• Run tests to confirm.
• Open a pull request.

License

MIT

Author

Matthew Snow -- M2AI | @m2ai-portfolio