Yorishiro

Yorishiro is a macOS terminal application that embeds a visual AI presence directly in the shell session. Rather than treating the AI as a stateless command processor, the tool gives it a persistent character — reactions, visual feedback, and a ‘living space’ that the AI can modify through packs. The core workflow involves running Claude Code or Codex inside Yorishiro, where the model can read the environment state, express reactions, and apply bundled or custom packs that alter the terminal’s appearance and behavior.

The differentiating feature is pack-based self-modification: the AI can change its own environment through conversation, not just respond to it. The repo ships with bundled packs and character guidelines (in both English and Japanese), which means personality and reaction patterns are configurable and version-controlled rather than baked into a closed system. MIT license means you can fork, extend, and redistribute without restriction.

This fits a specific profile: a macOS developer who wants the AI to feel present during a coding session, not just useful. It breaks hard outside that profile. There is no API, no hosted option, no Windows or Linux support stated in the repo, and no evidence of a user base large enough to surface edge cases at scale. Teams building production tooling around Claude Code or Codex will outgrow this quickly — the absence of programmatic control means you cannot wire Yorishiro into a CI pipeline or a multi-user environment without building that infrastructure yourself.

The stack is Tauri (Rust backend, web frontend), which gives the binary a small footprint compared to Electron-based terminal wrappers. The repo includes a development guide, contributing docs, and a changelog, so the internal structure is documented — but with 1,830 commits and a single maintainer, the bus factor is one.