LLM Council (with Installer)

LLM Council — ask one question to many models, let them critique each other, get a synthesized chairman answer.

This skill is the fastest way to run it: one command installs dependencies, configures credentials, and launches both backend and frontend. No manual setup, no API key prompts.

OpenClaw-native: Credentials resolve automatically from OpenClaw config or workspace .env. Falls back to the local OpenClaw gateway (port 18789) if no OpenRouter key is found.

Two Ways to Use LLM Council

| Mode | Best For | Command | |------|----------|---------| | Quick answer | Fast decisions, mobile, casual questions | /council "Your question" (requires ask-council skill) | | Full discussion | Deep research, exploring disagreements, seeing all model responses | /install-llm-council then open browser at :5173 |

Slash Command

/install-llm-council [--mode auto|dev|preview] [--dir PATH]

When the user says /install-llm-council, run:

bash ~/.openclaw/skills/install-llm-council/install.sh

The script will:

1. Resolve credentials — env var → workspace .env → OpenClaw local gateway (no prompt ever)
2. Clone or pull https://github.com/jeadland/llm-council to ~/workspace/llm-council
3. uv sync — Python backend dependencies
4. npm ci — frontend dependencies
5. Write .env — API key/URL for OpenRouter direct or OpenClaw gateway mode
6. Start app — uses hardened start.sh with mode-aware startup and health checks
7. Auto-handle port conflicts — selects safe fallback ports when defaults are busy
8. Print practical access URLs — Caddy route and common direct fallbacks

Flags

| Flag | Default | Description | |------|---------|-------------| | --mode auto | auto | Detect Caddy on :5173 and prefer preview mode; otherwise dev mode | | --mode dev | — | Run Vite dev server (hot reload, port 5173 default) | | --mode preview | — | Build + run Vite preview (port 4173 default) | | --dir PATH | ~/workspace/llm-council | Override clone directory |

Credential Resolution (OpenClaw-native)

The installer never prompts for API keys. It resolves credentials in this order:

1. Environment — OPENROUTER_API_KEY already exported
2. Workspace .env — ~/.openclaw/workspace/.env contains OPENROUTER_API_KEY=...
3. OpenClaw gateway — reads ~/.openclaw/openclaw.json → gateway.auth.token + gateway.port
   • Sets OPENROUTER_API_URL=http://127.0.0.1:<port>/v1/chat/completions in .env
   • Uses the gateway token as the bearer key (OpenAI-compatible endpoint)

Ports

| Service | Port | Notes | |---------|------|-------| | Backend (FastAPI) | 8001 | Always | | Frontend dev | 5173 | --mode dev (default) | | Frontend preview | 4173 | --mode preview |

Files

| File | Purpose | |------|---------| | SKILL.md | This file — skill documentation | | install.sh | Main one-shot installer/launcher | | stop.sh | Stop background services | | status.sh | Check if services are running | | pids | Saved PIDs for background processes |

Agent Instructions

When user says /install-llm-council or "install llm-council" or "start llm council":

bash ~/.openclaw/skills/install-llm-council/install.sh

Report back the access URL from the script output (e.g. http://10.0.1.X:5173).

To stop:

bash ~/.openclaw/skills/install-llm-council/stop.sh

To check status:

bash ~/.openclaw/skills/install-llm-council/status.sh

Example Output

✅ LLM Council installed and running!

  Mode:     dev
  API:      openrouter
  Backend:  http://127.0.0.1:8001
  Frontend: http://10.0.1.42:5173

  Stop:     bash ~/.openclaw/skills/install-llm-council/stop.sh
  Status:   bash ~/.openclaw/skills/install-llm-council/status.sh