CodeLeash

CodeLeash is an opinionated framework for full stack web development in a monorepo using Claude Code exclusively. It is a full-stack monorepo with its backend in python using FastAPI, database in Supabase, and frontend in React, built with Vite. Your production environment only requires uvicorn and does not require a frontend server, as all React assets are built and served statically. Commits can atomically change the frontend and backend as well as all types used to exchange data between them.

The stack:

- Claude Code
- Supabase
- FastAPI & python & pydantic & pytest
- React, TypeScript & Vite & vitest & Tailwind
- prek (a faster, Rust-based pre-commit alternative)
- pydantic2ts (for converting python types to TypeScript)
- Linters & code formatters: black, ruff, eslint, vulture...

As opposed to being "unleashed", this framework revolves around the concept of not trusting the agent to do the right thing, but forcing it to. Why spend tokens validating things that can be validated deterministically with a little bit of code. The framework includes tools and linters to validate rather than trust LLM audits. Codifying the checks helps make building with AI more predictable, cheaper and faster. It is also a never-ending process, continuously iterating to improve the framework. The more checks and guardrails are in place the more the LLM's work can be trusted and the more work can be done with more cursory code review. However - you should still be reviewing every line of code. This is not about vibecoding; it's about removing yourself from progressively more and more loops over time so you can gain leverage, but doing it at a low level instead of a high one, to create _more_ predictability in the process, so you can reason about it and skip over steps you'd normally…