codellm

codellm is an AI-powered coding agent built using the Google GenAI SDK. It leverages Large Language Models (specifically gemini-2.5-flash) to act as a thought partner and execution engine for coding tasks. The agent can explore a workspace, read file contents, write new code, and execute Python scripts to solve problems autonomously.

Features

• Autonomous Coding Agent: Acts as a helpful AI assistant that makes and executes plans to fix issues or fulfill user requests.
• File Management: Can list files/directories and read file contents within a restricted working directory.
• Code Execution: Executes Python scripts with optional arguments using a secure subprocess environment.
• File Writing: Can create or overwrite files in the local workspace.
• Tool Integration: Uses Google's Function Calling mechanism to map LLM intents to local Python functions.

Directory Structure

• main.py: The entry point of the application, handling the chat loop and function dispatching.
• prompts.py: Contains the system instructions that define the agent's behavior and constraints.
• functions/: Contains the core tool implementations:
  • get_files_info.py: Lists files and metadata.
  • get_file_content.py: Reads file text.
  • run_python_file.py: Executes Python scripts.
  • write_file.py: Writes data to disk.
  • call_function.py: A wrapper that routes LLM function calls to the appropriate Python code.

Prerequisites

• Python >= 3.12
• Google GenAI API Key (stored in a .env file)

Installation

1. Clone the repository:

   git clone <repository-url>
   cd codellm

2. Install dependencies: The project uses google-genai and python-dotenv.

   pip install google-genai==1.12.1 python-dotenv==1.1.0

3. Configure Environment: Create a .env file in the root directory and add your Gemini API key:

   GEMINI_API_KEY=your_api_key_here

Usage

Run the agent by providing a user prompt as a command-line argument:

python main.py "your coding request here"

options

--verbose: Enable this flag to see detailed logs of token usage and function call execution.

Security Note

The agent is constrained to a specific working directory (configured in functions/call_function.py as ./calculator) to prevent unauthorized file access or execution outside the intended project scope.