Canopy

Canopy is the visible layer of a bigger system. Behind every venture I run — Kaeru, Aoi-san, Bui, Portfolio — sits a crew of specialist AI agents. Canopy is where I put them in one room, hand them a single real decision, and watch them argue it out before I sign off. Overcooked-style isometric, but the stakes are my actual roadmap.

The whole app is one loop: convene → debate → decide → approve. You ask the panel a question, they deliberate live, and nothing is final until you — the Overseer — say so.

What's inside

🎛️ The situation room

A hand-built isometric room where the meeting happens. Each agent has its own sprite and seat; a status line tells you whether the panel is idle, connected, debating, or done. It reads like a game screen on purpose — the point is to watch the work, not read a transcript dump.

🧑‍🤝‍🧑 A cross-project panel

The roster isn't hardcoded. Canopy discovers agents by reading the .claude/agents/*.md files from a shared groundwork repo and from each venture (Kaeru HQ, Aoi-san, Bui, Portfolio…). Whoever I've actually defined as an agent shows up on the panel — so the crew debating a call is the same crew that does the real work.

⚖️ Bounded, mediated debate

One question, one round, a fixed cast — bounded so it can't spiral. Two styles: mediated, where a chair routes the turns, and peer-to-peer, where the agents respond to each other directly. The debate produces meeting minutes and a proposed decision, not just chatter.

✅ Human-in-the-loop approval

Nothing ships on the model's say-so. When the panel lands on a recommendation I can Approve, Hold, or send it back for another round. Approved calls are written to a dated decision file in the ledger — so there's a paper trail, not vibes.

📡 Streamed live

The debate arrives over Server-Sent Events, so it unfolds in the browser as it happens instead of landing as one finished wall of text.

🔌 Bring your own engine

Runs in mock mode by default — no network, no API keys, fully explorable. Flip to live and it can talk to Claude, ILMU (a Malaysian LLM), or Ollama for fully-local models.

How it's built

• Next.js (App Router) + TypeScript + Tailwind v4, on Bun.
• SSE over websockets — an /api/events route streams the debate through a small pub/sub bus (lib/bus.ts); /api/{start,approve,roster} handle the rest.
• Roster discovery — lib/agents.ts walks groundwork and each venture repo for .claude/agents/*.md, so the panel is generated from real agent definitions rather than a static list.
• The orchestrator — lib/orchestrator.ts runs exactly one bounded, mediated debate and emits each turn as it resolves.
• Auth without pasting keys — lib/auth.ts can reuse my Claude Code OAuth token straight from the macOS keychain, so live mode needs no secret in an env file.
• The room — components/CanopyView.tsx renders the isometric canvas client-side; there's a sprite gallery at /sprites.

Why I built it

Most multi-agent tools hide the work: you get an answer and a shrug. I wanted the opposite — to see the disagreement, keep the veto, and walk away with a written record I trust. Canopy makes the deliberation legible and keeps me the decision-maker, not the black box.