---
name: llm-council
description: >-
  Multi-model LLM Council with live dashboard. Query multiple AI models simultaneously,
  see responses side-by-side in a swarm-style dashboard, synthesize consensus, and run
  anonymous model-to-model voting. Use when the user asks to start the LLM council,
  compare models, query multiple models, convene council, ask all models, do model
  comparison, run multi-model queries, or launch the council dashboard. Supports
  Claude Sonnet 4.5, Claude Opus 4.5, GPT-4o, GPT-5.1, Gemini 2.5 Flash, and
  Gemini 2.5 Pro via AI Gateway.
---

# LLM Council

Query multiple AI models in parallel with a live web dashboard.

## Launch

```bash
cd ~/.claude/skills/llm-council/scripts
# Kill any existing server on the port
fuser -k 8787/tcp 2>/dev/null
# Start server
nohup python3 server.py > /tmp/council-server.log 2>&1 &
# Wait for startup, verify health
sleep 2 && curl -s --max-time 5 http://localhost:8787/health
# Export port for browser access
/app/export-port.sh 8787
```

Environment: `COUNCIL_PORT` (default 8787), `AI_GATEWAY_API_KEY` (required, auto-detected from environment).

## Files

- `scripts/server.py` - ThreadingHTTPServer, serves static files + API routes, SSE streaming
- `scripts/ai_gateway.py` - AI Gateway client: query, parallel query, streaming, synthesis, anonymous voting
- `scripts/static/index.html` - Dashboard UI (HappyCapy design system, light/dark theme)
- `scripts/static/app.js` - Client-side logic (model selector, SSE parsing, markdown rendering, voting UI)

## API

| Method | Path | Description |
|--------|------|-------------|
| GET | `/` | Dashboard |
| GET | `/api/models` | List models |
| GET | `/health` | Health check |
| POST | `/api/council/stream` | Query models (SSE) |
| POST | `/api/council/synthesize` | Synthesize consensus |
| POST | `/api/council/vote` | Anonymous model vote (SSE with keepalive) |

## Key Details

- Server uses `ThreadingHTTPServer` for concurrent request handling
- Vote endpoint streams SSE keepalive pings every 5s to prevent proxy timeouts (Cloudflare 524)
- Static files served with `Cache-Control: no-store, no-cache` headers
- Script tag uses cache-busting version param: `app.js?v=...` (bump when editing app.js)
- Model IDs use dots not hyphens: `anthropic/claude-sonnet-4.5` (not `claude-sonnet-4-5`)