Hi HN, I'm Thorsten. Caudra (
https://caudra.ai) is an open source TUI coding agent written in Rust. It started as a fork of Maki and has grown into its own project.
I spend most of my working day with coding agents and i have tried a lot of them. I kept adding more and more plugins to get the features I wanted. Most of the time though, i couldn't build the plugins i needed because I couldn't hook into the places I needed (e.g. proper /goal mode). It was also hard to keep them up-to-date, and I wanted one binary I could install in new environments and use right away. It also had to work with small/fast (even local) models as well as flagship ones.
I'd like to have feedback on these things:
- Steering: You can send a message as "next" (after the turn ends), "guide" (at the next safe point) or "replace" (right now). IMO, "next" is useful for simple ad-hoc chains.
- /btw: Ask a side question while the main agent keeps working. It reuses the prompt cache, so follow-ups are cheap. This works even during an active question prompt.
- Small/fast models: Caudra nudges the model when a turn stalls, which helps models like Qwen3.8-27B that sometimes stop early.
The conversation is always append-only, so the provider cache stays warm. From using it myself, I see about 97% cache hits with anthropic models and 94% with openai.
It's still in preview. I'd like to hear what breaks and what you'd want before you'd consider to switch.
Repo: https://github.com/caudra/caudra
Clips: https://caudra.ai/gallery/
I spend most of my working day with coding agents and i have tried a lot of them. I kept adding more and more plugins to get the features I wanted. Most of the time though, i couldn't build the plugins i needed because I couldn't hook into the places I needed (e.g. proper /goal mode). It was also hard to keep them up-to-date, and I wanted one binary I could install in new environments and use right away. It also had to work with small/fast (even local) models as well as flagship ones.
I'd like to have feedback on these things:
- Steering: You can send a message as "next" (after the turn ends), "guide" (at the next safe point) or "replace" (right now). IMO, "next" is useful for simple ad-hoc chains.
- /btw: Ask a side question while the main agent keeps working. It reuses the prompt cache, so follow-ups are cheap. This works even during an active question prompt.
- Small/fast models: Caudra nudges the model when a turn stalls, which helps models like Qwen3.8-27B that sometimes stop early.
The conversation is always append-only, so the provider cache stays warm. From using it myself, I see about 97% cache hits with anthropic models and 94% with openai.
It's still in preview. I'd like to hear what breaks and what you'd want before you'd consider to switch.
Repo: https://github.com/caudra/caudra
Clips: https://caudra.ai/gallery/