A minimal LLM powered coding agent

This agent takes the natural language input ("example - explain how does the calculator render results to the console?") and attempts to read, generate, edit and execute the code. I uses Gemini API for generation (function calling) and python subprocess module for execution.

Config

In config.py you can set

• MODEL_NAME - Name of the Gemini model
• BUFFER_SIZE_MAX - Maximum input size, currently set to 10,000 characters
• WORKING_DIR - The model will have access to this directory and its subdirectories
• MAX_ITER - Maximum number of iterations that we want model to loop through

Create a .env in root directory of this project and set GEMINI_API_KEY there.

Instructions

This project uses uv for package and environment management. You can install uv from here.

In root directory, run

uv venv && uv sync

Then,

uv run main.py "you query"

You can use optional --verbose argument to see the functions called and additional info like token count.

Note : There is a demo calculator project inside calculator directory for you to experiment with.

This project was built using instructions from boot.dev. The calculator source is taken from their course.