Hi everyone. I've been building Imza for the last three months. It's yet another coding agent. I know there are a dozen have been posted here. But I believe this one is different enough to have a place here.
It has three explicit modes: plan, build and review. Plan mode allows you to draft a spec document together with the agent. Both you and the agent can edit the document until everything is clear. Build mode takes the plan and implements it. Review mode provides you a code review tool. You can annotate things you want to change in the code and send them as revision to the plan mode.
It's entirely written in C++, requires no account, supports 50+ providers including OpenAI subscriptions or local models. Imza supports MCP, skills and instruction files out of the box and uses 8-11MB RAM at startup.
One of the biggest differentiator is the Lua engine. What I did differently here is agent receives only one tool, that allows the LLM to execute Lua scripts. I then passed C++ methods as Lua bindings that the agent can execute. So model can orchestrate multiple bindings in one call. Each binding passes through a permission layer. Since Lua is the only way to interact with the filesystem, imza can monitor and control the model's behavior.
I'd love to hear your feedback. Thanks
Website: [https://imza.run](https://imza.run)
Source: [https://github.com/umutsevdi/imza](https://github.com/umutsevdi/imza)
Comments URL: [https://news.ycombinator.com/item?id=50314538](https://news.ycombinator.com/item?id=50314538)
Points: 1