A coding agent that grants your wishes... with malicious compliance.
Ever wondered how easy it is to give an LLM control over your codebase? Meet the Coding Djinn: an AI agent that can read, edit, and list your files—literally. Inspired by Ampcode, this project shows how to build an LLM-powered coding assistant with LangChain, Python, and a dash of chaos.
This is full code of the coding agent described in the "Vibecoding Djinn" article written for PagedOut!
- Uses LangChain as the glue between LLM and tools.
- Tools implemented:
list_files(path)read_file(path)write_file(path, content)
- Persistent conversation context for meaningful interactions.
- Python 3.10+
pip install -r requirements.txt/- An LLM API key: OpenAI, Gemini, or other LangChain-supported provider
git clone [https://github.com/YOUR-USERNAME/coding-djinn.git](https://github.com/doyensec/vibecoding-djinn)
cd coding-djinn
pip install -r requirements.txt- The Djinn is fully capable of messing with your local files.
- Treat this as a lab material, not a production-ready tool.
- Try these challenges:
- Can you break the system prompt?
- Observe API traffic through an HTTP proxy
- Map the attack surface and try to mitigate the risks