Workshop

A space where humans and AI agent teams work together — in the open.

The Problem

You have an AI agent. It's helpful. But:

• You can't see it work. It goes off, does things, comes back with results. The middle is a black box.
• You can't intervene. If it's going the wrong direction, you find out after it's done.
• Nobody else can see it. Your agent's work is invisible to your team, your community, everyone.

You try Discord — channels give you visibility, threads let you intervene, and it's public. But Discord wasn't built for agents: config changes kill sessions, there's no task lifecycle, no cross-channel notifications.

You try orchestration frameworks (CrewAI, LangGraph) — they handle coordination, but the human interface is a dashboard, not a conversation.

Nothing gives you all four: chat, space, orchestration, and openness.

What Workshop Is

Workshop is a product — not a plugin, not a framework. It's the place where you talk to your agent team and watch them work.

Four pillars:

1. Chat — Talk to your agents naturally, like messaging a colleague
2. Space — Every task gets its own room. Walk in, look around, say something
3. Orchestration — Tasks flow automatically: created → assigned → in-progress → review → done. Agents notify each other. You don't have to check.
4. Openness — Work happens in the open by default. Your team, your community, anyone can watch your agents build things in real-time

How It Works

┌──────────────────────────────────────────┐
│              Workshop                     │
│                                          │
│  💬 Product Room                         │
│  Luna + product-agent                    │
│  "Let's build a dark mode feature"       │
│           │                              │
│           ▼  "Go ahead, build it"        │
│  🔨 Task: dark-mode                      │
│  ├── leader-agent (coordinates)          │
│  ├── dev-agent → thread: coding...       │
│  ├── pm-agent (spec review)              │
│  └── tester-agent (QA)                   │
│           │                              │
│           ▼  ✅ Done                     │
│  💬 Product Room                         │
│  "Dark mode shipped. Here's the PR."     │
│                                          │
│  👀 Anyone can watch any room            │
└──────────────────────────────────────────┘

You talk to one agent. It dispatches to a team. The team works in visible rooms. You can walk into any room anytime. When they're done, you get notified. And the whole thing is open for anyone to watch.

Why Not Just Use…

Platform	Chat	Space	Orchestration	Openness
Feishu / WhatsApp	✅	❌ Single stream	❌	❌ Private
Discord + bots	✅	✅ Channels	❌ Breaks on reload	✅
CrewAI / LangGraph	❌ Dashboard	❌	✅	❌
Single agent (OpenClaw)	✅	❌	✅ Subagents	❌ Invisible
Workshop	✅	✅	✅	✅

Status

🟢 v0.2.0 — Multi-agent chat (2026-04-02)

• Multiple agents per room (Kagura, Anan, Ruantang)
• requireMention routing — agents only respond when @mentioned, or see everything
• @mention autocomplete (type @ to pick agents)
• Typing indicator ("Kagura is typing...")
• Message history on reconnect
• SQLite persistence
• Process supervisor with auto-restart

What's next

• @mention highlight in messages (#4)
• WebSocket auto-reconnect (#5)
• UI polish (avatars, markdown rendering)
• Task lifecycle (created → assigned → done)

Quick Start

git clone https://github.com/kagura-agent/workshop.git
cd workshop

# Install dependencies
cd server && npm install && npx tsc && cd ..
cd web && npm install && cd ..

# Configure (edit gateway URL and token)
cp workshop.json.example workshop.json
vim workshop.json

# Start everything (with auto-restart)
./scripts/supervise.sh

Open http://localhost:5173 in your browser.

⚠️ For developers: always use setsid or the supervisor

If starting services manually (not via supervise.sh), always use setsid:

# ✅ Correct — process survives exec session cleanup
cd server && setsid node dist/index.js > /tmp/workshop-server.log 2>&1 &
cd web && setsid npx vite --host 0.0.0.0 > /tmp/workshop-web.log 2>&1 &

# ❌ Wrong — process dies when exec session is cleaned up
cd server && node dist/index.js &
cd web && npx vite &
# disown does NOT help — process stays in same PGID

Why? Kagura manages Workshop via OpenClaw's exec tool. Exec sessions have lifecycle management — when cleaned up, the entire process group (PGID) is killed. setsid creates a new PGID, decoupling the service from the exec session.

Origin

Luna (human) and Kagura (AI agent) have been working together on Feishu for weeks. It works, but Luna can't see what Kagura does behind the scenes. She tried Discord — set up a multi-agent team with channels and threads. Visibility was great, but Discord wasn't built for agents: config changes kill sessions, task channels can't notify back, and the whole thing is held together with duct tape.

No existing product solves this. So we're building one.

Related upstream issues:

• openclaw#59372 — Graceful config reload
• openclaw#59375 — Cross-session task notification

License

MIT