I've been working on Ossature, an open-source tool that takes a different approach to LLM code generation. Instead of prompting and hoping, you write specs describing what your software should do, Ossature breaks them into a build plan, and then executes it step by step with an LLM generating code under tight constraints.

The core idea: specs are the source of truth and the human stays in the loop. You review the plan before anything gets built. You can step through tasks one at a time or let it run and intervene on failures. When task 14 of 30 fails, you fix that task and keep going instead of starting over. When requirements change you update the spec and only the affected parts get rebuilt.

The workflow is validate -> audit -> build. Validate checks structure (no LLM). Audit sends specs to an LLM for review and generates the build plan. Build executes it task by task with verification after each step.

Works with Anthropic, OpenAI, Mistral, Google, etc. and local models through Ollama.

Python 3.14+, MIT licensed.

GitHub: https://github.com/ossature/ossature

Docs: https://docs.ossature.dev

Some examples: https://github.com/ossature/ossature-examples

This is v0.0.1. Right now it works best for greenfield projects. I want to support workflows against existing codebases but honestly I'm not sure yet what the right approach looks like there. If you have thoughts on that or anything else, open an issue or start a discussion on the repo.