The name is Japanese — a yorishiro is an object inhabited by a spirit. That’s roughly the idea: instead of treating the terminal as a window onto an agent, treat it as the place the agent lives.
The active agent appears as a 3D character. States such as thinking, errors, and completion are expressed through the character and its surroundings instead of being left buried in scrollback.
I started building it because long sessions with coding agents were wearing me out. Part of the problem, I think, was staring at a stream of text and constantly trying to infer what state the agent was in.
The AI characters I grew up with in fiction felt different. They expressed emotion through their bodies and facial expressions, and felt like beings with wills of their own. I began to think that this wasn’t merely decoration, but an interface question about how we perceive and understand AI.
A few things I ended up caring about:
* Reactions shouldn’t wait for the LLM.
A separate Reflex Layer watches raw PTY output. When a build fails, the character can react immediately, before the model has said anything.
Lighting, scenes, and effects also communicate what is happening with the agent.
Personas, reflexes, scenes, and effects are defined as JavaScript Packs. You can edit them yourself, or ask the agent to change its own appearance and surroundings. Changes hot-reload without interrupting the session and can be rolled back with one click.
> brew install --cask sktkkoo/yorishiro/yorishiro
Honestly, I still don’t know whether this makes long sessions easier, or whether it just becomes distracting once the novelty wears off.
If you try it, I’d like to hear where it starts to feel annoying.
chriswunan•20h ago