Tell Yogen what's on your mind — a startup idea, a price change, a decision you keep putting off. It spins up 12 named AI experts and a 488-agent crowd, they argue about it for six rounds, opinions cascade through a social graph, and you get a verdict with a full written report.
Not a chatbot giving you one opinion. A simulated debate where positions are earned.
Try it hosted (3 free sims): www.yogen.uk · Full demo video
git clone https://github.com/michaelwrites67-ctrl/yogen.git
cd yogen
npm install
cp .env.example .env # add your Anthropic API key
npm start # → http://localhost:3005That's it. Self-hosted Yogen has no simulation limits — you pay only your own Anthropic API usage (a full 500-agent sim costs roughly $0.20).
Requirements: Node 18+, an Anthropic API key.
your idea
│
▼
┌─────────────────────────────────────────────────────┐
│ 1. Market intel + knowledge-graph extraction │
│ (from your idea + any document you attach) │
├─────────────────────────────────────────────────────┤
│ 2. 12 named experts generated for YOUR domain │
│ 4 positive · 4 sceptical · 4 negative │
├─────────────────────────────────────────────────────┤
│ 3. 488 background agents in a social graph, │
│ starting perfectly split 33/33/33 │
├─────────────────────────────────────────────────────┤
│ 4. Six rounds of debate. Experts argue; stances │
│ change only when a point can't be answered. │
│ Opinion cascades ripple through the crowd. │
├─────────────────────────────────────────────────────┤
│ 5. Verdict + prediction report: strongest signals, │
│ the one thing to check, concrete next steps — │
│ each doable solo, in a browser, under 30 min. │
└─────────────────────────────────────────────────────┘
You can argue back. Mid-debate, type into "Speak to the swarm" — correct an agent's assumption, add evidence — and the debate absorbs it in the next round.
Every simulation ends with an Oracle Verdict, the final swarm split, and a full written report — strongest signals for and against, the single most important unknown, and next steps you can each do solo, in a browser, in under 30 minutes.
- Balanced start. The swarm begins exactly 33% positive / 33% sceptical / 33% negative. The verdict must be earned by the debate, never baked in.
- Stance discipline. Agents hold their position unless a peer lands a point they cannot answer. Finding fault is not treated as insight.
- Conviction score. Every verdict ships with a quality signal — how strongly the swarm converged and how many agents genuinely changed their mind — so you know whether to trust it.
- No invented facts in reports. Agents improvise scenarios while debating (that's what debaters do); the final report is constrained to what you actually said, and tells you what to go and verify.
Plain Node.js + Express, one HTML file per page, zero front-end framework. Agents run on Claude Haiku (fast, cheap); report synthesis on Claude Sonnet. File-backed JSON storage — no database to set up.
server.js routes, auth, SSE streaming, quotas
simulation.js the debate engine — agent generation, rounds, report
rule-engine.js 488-agent crowd: social graph, cascades, coalitions
agents.js expert persona generation
public/app.html the simulator UI
Everything is optional except the API key — see .env.example. Set UNLIMITED_SIMS=false and DAILY_SIM_LIMIT if you deploy it publicly and want quotas; Stripe and Google-OAuth vars exist for hosted billing/login and can be ignored.
AGPL-3.0. Use it, fork it, learn from it, self-host it freely. If you run a modified version as a service, you must open-source your modifications. The hosted service at yogen.uk is operated by Runn Ltd.