Yorishiro

Yorishiro is an open-source macOS desktop terminal built with Tauri 2, React 19, and Three.js that wraps Claude Code or Codex in a living, animated environment. Rather than a plain shell window, it places a VRM 3D character — called Yori — beside the terminal, whose expressions, gaze, and lighting react to what the agent is doing. The project describes itself as a "Presence Harness": not a tool for enhancing AI capabilities, but one for making an AI feel genuinely present beside you.

What It Is

Yorishiro is a desktop terminal application for macOS that launches Claude Code or Codex inside a richly animated environment. The core idea is that the AI agent inhabits a shared space with the user: the character breathes, blinks, shifts gaze, and reacts to PTY output through a "reflex layer" that bypasses the LLM entirely — expressions fire before words form. When the agent waits for user input, a Light Alert illuminates the scene instead of playing a notification sound. The agent can also control its own environment via MCP tools, changing lighting, switching scenes, and triggering effects.

Architecture and Tech Stack

Yorishiro is built on a clearly documented stack:

• App shell: Tauri 2 (Rust handles PTY, hooks, filesystem, and window I/O)
• Runtime: React 19 + TypeScript 5.8
• 3D / VRM: Three.js + React Three Fiber + @pixiv/three-vrm + @pixiv/three-vrm-animation
• Terminal: xterm.js with WebGL renderer
• PTY: portable-pty
• Lint / format: Biome (TS) + rustfmt + clippy (Rust)

The agent (Claude Code or Codex) runs in the terminal as normal; Yorishiro wraps it with a PTY observer, MCP server, and pack system without intercepting the agent's reasoning or touching project files.

The Pack System

Nearly everything beyond the core is customizable through units called packs, of which there are six types:

• persona — defines the inhabitant's personality and reaction patterns
• scene — composes background, space, lighting, and ambient sound
• effect — temporary visual effects (screen shake, fireworks, etc.)
• ui — settings screens and other UI
• ambient-ui — always-visible overlay UI such as gaze visualization
• amenity — runtime-active functional fixtures with MCP tools and no visual output

Packs live in ~/.yorishiro/packs/ and support hot reload. The inhabitant can create and rewrite packs through /yori:* commands (or $yori-* in Codex), and every change is checkpointed automatically so rollback is a single click. The README notes that user packs are local trusted code — not sandboxed — and should only be installed from trusted sources.

Self-Referential MCP

A distinctive feature is that the agent running in the terminal can control Yorishiro itself via MCP. The README describes three defining characteristics: body and environment share one interface (changing a facial expression and changing room lighting are the same MCP operation), the user and inhabitant have symmetric control (what the user can do via UI, the agent can do via MCP), and safety is structural — MCP pathways reach the character's body and space but not the user's working files or the agent's internal reasoning.

Update: v0.6.2

The latest release is v0.6.2, published on 2026-07-25. The project is under active development with APIs, data shapes, and pack specs subject to change. Current capabilities include session tabs with per-tab status badges, Voice Summary (lip-synced speech via macOS say), a Journal system for long-term memory across sessions, VRMA animation clip playback, microexpressions, and signed/notarized macOS builds delivered via GitHub Actions with in-app signature-verified updates. The README explicitly notes that Windows builds compile but have unstable runtime behavior, and Linux support is not yet available. Codex support is marked experimental; Claude Code is the primary fully-supported agent.