A transparent, hackable AI coding agent built in under 1200 lines
Most AI coding tools are black boxes. You send your code to a server, magic happens, and you get results back. But what exactly is happening? What data is being collected? What prompts are being used? How does the agent actually work under the hood?
This opacity creates several problems: privacy concerns, vendor lock-in, inability to customize behavior, and most importantly - you can't learn from or improve the system.
We set out to build a complete AI coding agent that anyone could understand, modify, and extend - all in under 1200 lines of code. Not 1200 lines of business logic plus thousands of dependencies. 1200 lines total.
This constraint forces clarity. Every line matters. Every feature is essential. The result is a system you can read in an afternoon, understand completely, and adapt to your needs.
$ find . -type f -name "*.py" -not -path "*/\.*" | xargs wc -l | tail -1
1187 totalEvery prompt, every decision, every line of code is visible and auditable.
Want to add support for a new language or tool? The code is designed to be modified.
Runs locally. Your code never leaves your machine unless you want it to.
No vendor lock-in. No API keys required. Modify, distribute, and use however you want.
Despite the constraint, Muonry provides a complete AI coding experience:
Planning, execution, and error handling
Read, write, create, and modify files
Run tests, install packages, execute commands
Research APIs, documentation, and solutions
Understand existing codebases and dependencies
Natural language interaction and feedback