LLM Council

The LLM Council feature lets you query multiple AI models simultaneously and combine their responses into a single, higher-quality answer. Instead of relying on one model, you assemble a council of 2 or more models that deliberate together.

What Is a Council?

A council is a group of AI model profiles that work together on each message you send. Each council member can be a different provider, model, and system prompt. When you send a message, all members respond, optionally review each other's work, and a designated chairman synthesizes the final answer.

Councils are especially powerful for:

• Research -- get multiple perspectives on a topic
• Code review -- have different models check each other's work
• Decision-making -- weigh competing viewpoints
• Quality assurance -- catch errors that a single model might miss

7 Council Styles

Each council style defines how the models interact. Choose the style that best fits your use case.

Council (Default)

The standard 3-phase deliberation pipeline:

1. Phase 1 -- Fan-Out: Your message is sent to all members in parallel. Each model generates its own independent response.
2. Phase 2 -- Peer Review: Each member reviews the other members' responses and ranks them from best to worst with reasoning.
3. Phase 3 -- Chairman Synthesis: A designated chairman model reads all responses and peer reviews, then produces a comprehensive final report.

Best for: Research, analysis, complex questions where you want thorough coverage.

Compare

All models respond in parallel, and their responses are displayed side-by-side in a grid. There is no voting, no peer review, and no synthesis step.

• Phase 1 only (fan-out)
• Responses shown in columns
• No chairman, no final answer

Best for: Comparing model capabilities, testing prompts across models, seeing how different models approach the same question.

Arena

Models compete head-to-head. All models respond, then peer review determines a winner. The best response wins.

Best for: Finding the single best answer, competitive benchmarking.

MoA (Mixture of Agents)

Layered refinement where each model builds on the previous model's output. Instead of independent parallel responses, models iteratively improve on each other's work.

Best for: Tasks that benefit from iterative improvement, like writing and editing.

Router

Smart routing that picks the best model for each query automatically. Instead of querying all models, the router analyzes your message and sends it to the single most appropriate model.

Best for: Cost optimization, routing different types of questions to specialized models.

Debate

Models are assigned opposing sides (FOR and AGAINST) and argue their positions:

1. Opening Arguments: Each model argues its assigned side
2. Rebuttals: Models respond to the opposing side's arguments (configurable number of rounds)
3. Moderator Verdict: The chairman analyzes both sides fairly and delivers a verdict

Debate sides are auto-assigned (alternating) or can be set manually per member.

Best for: Exploring controversial topics, stress-testing ideas, finding weaknesses in arguments.

تلميح

You can set the number of debate rounds in the council settings. More rounds means deeper argumentation but higher API costs.

Consensus

Models vote on the best answer without a synthesis step:

1. All models respond independently (fan-out)
2. Each model reviews and ranks the other responses (peer review)
3. The response with the highest vote score is selected as the final answer

No chairman synthesis -- the winning response is used as-is.

Best for: When you want the crowd's pick rather than a synthesized summary.

Creating a Council

Using the Wizard (New Users)

1. When the wizard appears, choose Cluster at Step 0
2. Select models from the presets tab or add custom models
3. Connect API keys for each provider (skipped for free models)
4. Configure council settings: name, icon, style, and member summary
5. Click Ready to create the council

From Settings (Existing Users)

1. Open Settings and go to the Profile tab
2. Add members using the profile picker (each member references a saved profile)
3. Set the council style from the style grid
4. Configure the chairman, voting mode, and number of rounds
5. Save the profile

معلومات

Council members are always profile references. Create individual profiles first (one per model), then assemble them into a council. This lets you reuse the same profile across multiple councils.

Council Members

Each council member has its own settings that can override the council defaults:

Setting	Description
Profile	Which saved profile (provider + model) to use
System Prompt	Override the member's default system prompt
Temperature	Override temperature for this member
Max Tokens	Override max output tokens
Reasoning Effort	Set thinking/reasoning level (Off, Low, Medium, High, Highest)
Debate Side	For debate style: assign For, Against, or Auto

Members are labeled A, B, C, etc. for identification in the council output.

Chairman Role

The chairman is the model responsible for synthesizing the final answer in council, arena, and debate styles. By default, the first member (A) is the chairman, but you can change this in the council settings.

The chairman receives:

• The original user message
• All member responses from Phase 1
• All peer review rankings from Phase 2
• Instructions to produce a comprehensive research report

تلميح

Choose your most capable model as chairman. The chairman does the heaviest lifting -- it needs to process all other responses and produce a coherent synthesis.

Voting Modes

When peer review is enabled, members rank each other's responses. The voting mode determines how those rankings are tallied:

Mode	How It Works
Weighted	Members earn points based on ranking position. First place gets N points, second gets N-1, etc.
Plurality	Only first-place votes count. The response with the most first-place rankings wins.

Vote scores are displayed in the final output next to each member's response.

Cost Estimation

Running a council multiplies API usage by the number of members and phases. Before sending a message, the platform estimates the cost based on:

• Number of members
• Number of phases (varies by style)
• Expected token counts
• Per-model pricing from the registry

The cost estimate is shown in the council output footer after each response:

API calls: 7 | Tokens: 24,531 | Est. cost: $0.1847

warning

Councils with many members and paid models can be expensive. Compare mode is the cheapest (Phase 1 only), while full council or debate with multiple rounds is the most expensive.

Live Streaming

Council deliberation streams in real time. During Phase 1, you see each member's response appear in a grid as it generates. Status indicators show which members are thinking, streaming, done, or failed.

During Phase 3, the chairman's synthesis streams token-by-token just like a regular chat response.

Example Use Cases

Style	Use Case	Example Members
Council	Research report on a technical topic	Claude (analytical) + GPT-4o (broad) + Grok (contrarian)
Compare	Testing a prompt across models	Gemini Flash + Claude Haiku + GPT-4o mini
Arena	Finding the best code solution	Claude Sonnet + GPT-4o + DeepSeek Coder
MoA	Polishing a blog post	GPT-4o (draft) + Claude (edit) + Gemini (polish)
Router	Mixed daily use	Math model + Code model + Creative model
Debate	Should we use microservices?	2 models FOR + 2 models AGAINST
Consensus	Which framework to use?	3-5 diverse models voting

Free Model Councils

You can build councils entirely from free models (OpenRouter free tier, Gemini free tier). The platform automatically handles rate limiting for free models by sending requests sequentially instead of in parallel.

ملاحظة

Free models have lower rate limits (typically 8 requests per minute). Sequential execution means council deliberation takes longer, but it works reliably without hitting rate limits.