Show HN: Caudra – An open-source terminal coding agent in one native binary (github.com)

Hi all, Thorsten from Caudra (https://caudra.ai) here. I'm building Caudra, an open source coding agent written in Rust with a good, albeit opinionated, experience out-of-the-box. You don't need mess with plugins or keep them up-to-date.

As a software engineer/data architect, I spend most of my day with coding agents and i tried many of them. I needed something that combines all the great features of these agents + new features in a new standalone TUI binary that i can easily get up running in new environments. I also needed something that works great with small models (including local ones) as well as with flagship models. Caudra does a lot of hand-holding and context management automatically so agents stay on track.

QoL improvements:

It supports eager streaming so you don't need to wait on a spinning icon and directly see how tool calls are being formed. Subagents are interactive as soon as possible (already after the first instruction token is being streamed). Define goals (like in claude code) to make sure that the agent actually works toward it without going astray. Continual learning: a background agent that automatically remembers important details and quirks in your setup, the way you desire things, and saves them in your AGENTS.md, AGENTS.local.md, global AGENTS.md, and memory system. Three message steering types: "next" waits until user turn, "guide" sends at the next safe boundary (like most coding agents), "replace" inserts now. "next" is really cool because it only sends the message once the turn is completely over (allows some really rudimentary ad-hoc chains). Autopilot can answer permissions prompts and questions for you and makes sure the agent is not stalled before reaching (or failing) a goal. Caudra nudges the model: stalled turns (particularly useful for smaller models like Qwen-3.8-27b that sometimes just stop at turns), nudge to use question tool when models ends a turn which could more easily be handled by just answering question prompt (requires decision model to be configured). Integrates well with Herdr and Orca you might be already using, proper support for git worktrees, (git-based) automatic workspace snapshots. Markdown rendering including math and mermaid charts (common ones). /btw lets you always ask side questions (including follow-ups with almost 100% cache hits) without interrupting the main agent (can even be used during an active question tool prompt to ask for more context).

Token savings:

Even though many things can be hot (re)loaded, they never break the provider prompt cache (conversations are append only). This results in pretty good provider cache rates: anthropic models ~97% (98% with cache warming) and openai models ~94%. Background (async) subagents and tasks turn Caudra into an event driven coding agent. This results in faster throughput and more provider cache hits (caches are still warm). Deferred questions let you answer some questions at your convenience without blocking the main agent. Code mode runs Monty Python in a separate process: all configured MCP resources are exposed here as well as some of Caudra's internal tools that shouldn't always stay armed in system context. Efficient (infinite-like) memory system (with good recalls), RTK-like token reductions (from shell outputs), file index and source code context/refs tools (similarly to ripwire) to efficiently navigate codebases.

Permissions and security:

Decomposes complex commands e.g. `make lint && make test` are really two different commands (can learn and detect permission patterns as well). Built-in decision engine (JEV and OpenAI decision compatible) for content screening, shell effects, re-ranking, etc. Only forwards safe or whitelisted environment variables to shell tool. By default doesn't let you fetch loopback or LAN IPs unless you allow it.

This was really just the tip of the iceberg. Have fun trying it out.