Codex, Claude Code, OpenCode, Kilo Code, Kimi CLI, Grok Build, OpenClaw, Hermes Agent.
All useful. Still not enough for the kind of harness I want.
I need something that remembers what I worked on in a folder last week without replaying the whole session. Something I can trust when agents are touching real code.
So I’m building Yemrecode.
It started as an OpenCode fork, but a lot of it is now heavily customized or rebuilt from scratch. No TUI. Just a Linux-native GUI.
The first version is almost complete and already works smoothly in most places.
A few core ideas:
I can select multiple LLMs at once, and Yemrecode treats them as one combined entity.
I can set a /yunus-goal, then agents manage the work around it. Some build. Some demand proof.
I can run swarms of different LLMs toward the same goal, with each one able to run its own swarm too.
YOLO mode can be toggled anytime, but agents still cannot modify Yemrecode itself.
There is also an anti-AI-slop controller that catches lazy AI-style UI/UX before it enters the code and forces the responsible agent to redo it.
The path detector does the same kind of protection for file writes, checking intent before allowing agents to touch suspicious locations.
And when a folder passes around 1,000 lines of code, Yemrecode automatically initializes git. At that point, it is probably not just throwaway code anymore.
I’m not publishing it yet. I want to use it hard for a few months first.
After that, maybe.