Council — Multi-Model Advisory Skill for OpenClaw

A simple OpenClaw skill that gives you a multi-model AI council. Ask a question, get independent analyses from multiple AI models, and receive a synthesized response showing where they agree, disagree, and what they recommend.

Works with your existing subscriptions — no API keys required. Uses OpenClaw's OAuth (ChatGPT Plus) and setup-token (Claude Pro) authentication.

What It Does

When you say something like "Ask the council: should I form an LLC or sole proprietorship?", the skill:

1. Your primary model (e.g., Claude) forms its own analysis independently
2. Sub-agents spawn on your other models (e.g., GPT, Gemini) for separate analyses
3. All analyses are synthesized into a structured response:

## Council Response

### Consensus
Where all models agree.

### Model A's Analysis
Unique insights from your primary model.

### Model B's Analysis
Unique insights from the second model.

### Model C's Analysis (if configured)
Unique insights from the third model.

### Divergence
Where they disagree and why it matters.

### Recommendation
Synthesized recommendation weighing all perspectives.

4. The full response is automatically saved as a searchable markdown file

Why This Exists

Existing multi-model solutions like LLM Council require API keys, which means paying per-token on top of subscriptions you already have.

This skill works with web subscriptions you already pay for:

• ChatGPT Plus ($20/mo) via OpenClaw's OAuth
• Claude Pro ($20/mo) via setup-token
• Any other provider OpenClaw supports

No extra costs. No API key management.

Installation

1. Copy the skill folder

cp -r skills/council ~/.openclaw/workspace/skills/council

2. Create the archive directory

mkdir -p ~/.openclaw/workspace/council

3. Restart the gateway

systemctl --user restart openclaw-gateway

4. Verify

openclaw skills list

You should see council listed as ready.

Configuration

Two models (default)

The skill ships with a two-model council. Edit SKILL.md Step 4 to set your second model:

Primary Model	Second Model (set in SKILL.md)
anthropic/claude-opus-4-6	openai-codex/gpt-5.3-codex
anthropic/claude-sonnet-4.5	openai/gpt-4o
openai/gpt-4o	anthropic/claude-opus-4-6

Three or more models

To add more council members, duplicate the sessions_spawn block in Step 4 of SKILL.md for each additional model:

sessions_spawn(
  task: "[same prompt]",
  label: "Council Member C",
  model: "google/gemini-2.5-pro",
  notify: false
)

Each model runs independently and in parallel. The primary model synthesizes all responses in Step 5. Adjust the response format in Step 5 to include a section for each council member.

All models must be authenticated in OpenClaw. See OpenClaw docs for setup.

Trigger Phrases

The council activates on natural language — no slash commands or mode switches needed:

• "Ask the council: ..."
• "Through the council, ..."
• "Council mode — ..."
• "I want all models to look at ..."
• "Get a second opinion on ..."
• "What do both think about ..."

Normal messages without these phrases are unaffected.

Saved Responses

Every council response is automatically saved to:

~/.openclaw/workspace/council/YYYY-MM-DD-HHMM-topic-slug.md

Example: council/2026-02-16-1200-massachusetts-llc-formation.md

Each file contains the original question, date, and the full structured council response.

Design Decisions

• Independent analysis first: The primary model forms its own analysis before seeing the other models' responses. This prevents anchoring bias.
• Parallel sub-agents: All council members analyze simultaneously — adding more models doesn't multiply wait time.
• Structured output: Consensus, individual analyses, divergence, and recommendation — not just answers pasted together.
• Keyword triggers: Works in normal conversation flow. No mode switching, no slash commands.
• Auto-archive: Every council session is saved for future reference and searchability.
• Subscription-friendly: Designed for users who have ChatGPT Plus and Claude Pro, not API keys.
• Extensible: Add as many models as you want by duplicating one sessions_spawn block.

Requirements

• OpenClaw (tested on v2026.2.14)
• Two or more AI providers configured and authenticated
• subagents.maxConcurrent >= number of council members in openclaw.json (default is 8)

License

MIT

Credits

Built by Hiro Fukushima.

Inspired by LLM Council by Andrej Karpathy and the OpenClaw skills ecosystem (council-of-the-wise, multi-viewpoint-debates).