GRAMCHECK

A Python tool to check text for grammar mistakes using local LLMs. It connects to any OpenAI-compatible server (e.g. llama-server) and streams corrections back as they are generated.

Because of the probabilistic nature of LLMs, results may or may not be correct. Take them with a grain of salt. Results are generated using a predefined seed to keep them consistent.

Currently does not support API keys. Future versions will plan API keys to enable support for online providers. Example tests are also planned for future versions.

Requirements

• Python 3.8+
• A running llama-server (or any OpenAI-compatible endpoint)
• A capable instruction-tuned model. Recommended: Q4 or higher quantization. Q2 quants and sub-4B models often fail to follow the output format.

Install

1. Install from PyPI: pipx install gramcheck.
2. Point it at your server: gramcheck --set-server-url http://127.0.0.1:8080 (writes _GC_SERVER_URL to .env in the project directory).

Usage

1. Check a file: gramcheck example.txt.
2. Check a single text: gramcheck -t 'Your text here'.
3. Check the Nth line in a file: gramcheck example.txt -n 0 (0-based, so -n 0 is the first line).
4. Check a file as a whole: gramcheck example.txt -a.
5. Show help: gramcheck --help.

Output

Each text is printed in red, followed by the model's corrections in green. Corrections are one line per error in the form original -> corrected. If a text has no errors, the model outputs OK.

Development

1. Install the uv project manager.
2. Clone the repository: git clone https://github.com/degD/gramcheck and run uv sync.
3. Point it at your server: gramcheck --set-server-url http://127.0.0.1:8080 (or edit .env directly).
4. Run locally with python gramcheck.py (or ./gramcheck on Unix-like systems).

Supported servers

Any endpoint that implements the OpenAI Chat Completions API (/v1/chat/completions) with streaming SSE. Tested with llama-server from llama.cpp.

See the documentation for llama-server-specific configuration.

License

MIT