Harness

---

An app I built for myself: dictation, chat without the persona, notes, and tasks. Offline-first. Full control over the inputs, the outputs, and everything in between. Commercial tools optimize for a mainstream audience and lock-in — this one is personal, portable, and legible, built to understand agentic systems by making one.

Principles

• Personal — opinionated and unfinished on purpose
• Portable — leave, switch models, keep your data
• Legible — see context assembly, system prompts, and tools per message
• Exploratory — a learning lab, not a product roadmap theater

Run

npm install
npm run dev

Needs Rust. On macOS, Xcode Command Line Tools for the speech helpers.

To iterate on the chat UI in a regular browser (Vite + localStorage, no Tauri/Rust):

npm run dev:browser

Opens http://localhost:5173. Chat works without a key (dummy streamed replies). Paste an OpenAI API key in System → Data, or set VITE_OPENAI_API_KEY in .env, for real completions. Conversations stay in this browser's localStorage — not the desktop profile. Dictation, sync, tools, and images are stubbed.

npm run dist:mac          # signed DMG — see BUILD.md
npm run demo              # watch the site's demo scenes live (player at /demo-player.html)
npm run capture:site      # render every scene into site/assets/ (PNG stills, MP4 clips)

Site screenshots and clips come from demo scenes in src/renderer/demo/scenes.ts: the real app on a scripted backend (demoAdapter.ts) with seeded data, a drawn cursor, scripted replies, and simulated Fn dictation. npm run demo plays them with hot reload; capture:site steps each one frame by frame on a paused clock in headless Chromium, so takes are identical and need no permissions. --only <scene> renders one; --gif also writes GIFs.

Surfaces

Desktop	Tauri (Rust + React) — chat modes (Chat/Decide/Write/Refine), tools, memory, notes, tasks, dictation, R2 sync
Browser	Same React chat UI via npm run dev:browser — localStorage; dummy replies unless an API key is set; debug shell, not a product client
Mobile	Native iOS companion — chat, capture, sync QR pairing (ios/README.md)
Models	OpenAI API or OpenAI-compatible locals (e.g. Ollama)
Speech	Apple Speech on macOS and iOS

Docs

• ROADMAP.md — outcomes and what’s active
• BUILD.md — packaging, signing, Dev vs installed
• SECURITY.md — trust boundaries and reporting
• docs/glossary.md — shared product vocabulary
• AGENTS.md — notes for coding agents (contracts, IPC, parallel-work board)
• CLAUDE.md — Claude Code entrypoint (@AGENTS.md)