Conclave MCP

An MCP (Model Context Protocol) server that provides access to a "conclave" of LLM models, enabling any MCP-compatible client to consult multiple frontier models for diverse opinions, peer-ranked evaluations, and synthesized answers.

Why This Exists

When working with an AI assistant, you're getting one model's perspective. Sometimes that's exactly what you need. But for important decisions—technical architecture, business strategy, creative direction, complex analysis, or any situation where blind spots matter—a plurality of opinions surfaces alternatives you might miss.

Conclave brings democratic AI consensus to any workflow.

Instead of manually querying multiple AI services, you can consult the conclave through Claude Desktop, Claude Code, or any MCP client. Get ranked opinions from multiple frontier models (GPT, Claude, Gemini, Grok, DeepSeek) and receive a synthesized answer representing collective AI wisdom.

Use cases include:

Technical: Architecture decisions, code review, debugging, API design
Business: Strategy analysis, proposal review, market research synthesis
Creative: Writing feedback, brainstorming, editorial perspectives
Research: Literature review, fact-checking, multi-perspective analysis
Decision-making: Pros/cons analysis, risk assessment, option evaluation

Inspired by Andrej Karpathy's llm-council concept. This project reimplements the core ideas as an MCP server for seamless integration with AI-assisted workflows.

How It Works

The conclave operates in up to 3 stages:

┌─────────────────────────────────────────────────────────────────┐
│  Stage 1: OPINIONS                                              │
│  Query multiple LLMs in parallel for independent responses      │
│  (GPT, Claude, Gemini, Grok, DeepSeek, etc.)                   │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Stage 2: PEER RANKING                                          │
│  Each model anonymously evaluates and ranks all responses       │
│  Aggregate scores reveal best performers (lower = better)       │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Stage 3: SYNTHESIS                                             │
│  Chairman model synthesizes final answer from collective wisdom │
│  Consensus level reported (strong/moderate/weak/split)          │
│  Tiebreaker vote cast if conclave is split                      │
└─────────────────────────────────────────────────────────────────┘

Features

Tiered queries: Choose cost/depth tradeoff (quick | ranked | full)
Three council tiers: Premium (frontier), Standard (balanced), Budget (fast/cheap)
Consensus protocol: Detects agreement level, triggers tiebreaker on splits
Odd conclave size: Ensures tiebreaker votes can break deadlocks
Rotating chairmanship: Weekly rotation prevents single-model bias
Chairman presets: Context-aware chairman selection (code, creative, reasoning)
Cost estimation: Know what you'll spend before querying
Eval-light: Standalone benchmark runner for tracking performance over time

Installation

Prerequisites

Get an OpenRouter API key from https://openrouter.ai/keys
Add credits to your OpenRouter account (pay-as-you-go)

Setup

# Clone the repository
git clone https://github.com/stephenpeters/conclave-mcp.git
cd conclave-mcp

# Create .env file with your API key
echo "OPENROUTER_API_KEY=sk-or-v1-your-key-here" > .env

# Install dependencies
uv sync

Configure Claude Desktop

Option 1: Desktop Extensions (Recommended)

Open Claude Desktop
Go to Settings > Extensions > Advanced settings > Install Extension...
Navigate to the conclave-mcp directory
Follow prompts to configure your OPENROUTER_API_KEY
Restart Claude Desktop

Option 2: Manual Config

Open Claude Desktop, go to Settings > Developer > Edit Config, and add the following to claude_desktop_config.json:

{
  "mcpServers": {
    "conclave": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/conclave-mcp", "python", "server.py"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-v1-your-key-here"
      }
    }
  }
}

Replace /path/to/conclave-mcp with your actual path, save, and restart Claude Desktop.

Configure Claude Code

Add the server using the CLI:

claude mcp add --transport stdio conclave -- uv run --directory /path/to/conclave-mcp python server.py --env OPENROUTER_API_KEY=sk-or-v1-your-key-here

Or copy .mcp.json.example to .mcp.json and update paths:

cp .mcp.json.example .mcp.json
# Edit .mcp.json with your paths and API key

Verify with /mcp in Claude Code or claude mcp list in terminal.

Available Tools

`conclave_quick`

Fast parallel opinions (Stage 1 only). Queries all conclave models and returns individual responses.

Cost: ~$0.01-0.03 per query

Use for: Quick brainstorming, getting diverse perspectives fast

`conclave_ranked`

Opinions with peer rankings (Stage 1 + 2). Shows which model performed best on this specific question.

Cost: ~$0.05-0.10 per query

Use for: Code review, comparing approaches, seeing which model "won"

`conclave_full`

Complete conclave with synthesis (all 3 stages). Includes consensus detection and chairman tiebreaker.

Cost: ~$0.10-0.20 per query

Options:

tier: Model tier - "premium", "standard" (default), "budget"
chairman: Override chairman model (e.g., "anthropic/claude-sonnet-4")
chairman_preset: Use a preset ("code", "creative", "reasoning", "concise", "balanced")

Use for: Important decisions, architecture choices, complex debugging

`conclave_config`

View current configuration: conclave members, chairman rotation status, consensus thresholds.

`conclave_estimate`

Estimate costs before running a query.

`conclave_models`

List all available models with selection numbers. Shows models grouped by tier with stable numbering:

Premium tier: 1-10
Standard tier: 11-20
Budget tier: 21-30
Chairman pool: 31-40

`conclave_select`

Create a custom conclave from model numbers. The first model becomes the chairman.

conclave_select(models="31,1,11,21")

Creates:

Chairman: #31 (deepseek-r1)
Members: #1 (claude-opus-4.5), #11 (claude-sonnet-4.5), #21 (gemini-2.5-flash)

Custom selection persists until server restart or conclave_reset.

`conclave_reset`

Clear custom conclave selection and return to tier-based configuration.

Custom Model Selection

For full control over which models participate in the conclave:

List available models: Use conclave_models to see all models with their numbers
Select your lineup: Use conclave_select(models="31,1,11,21") - first number is chairman
Query: Use conclave_quick, conclave_ranked, or conclave_full as normal
Reset: Use conclave_reset to return to tier-based config

Example workflow:

> conclave_models
## Available Models
### Premium Tier (1-10)
   1. anthropic/claude-opus-4.5
   2. google/gemini-3-pro-preview
   ...

> conclave_select(models="31,1,12,21")
## Custom Conclave Created
Chairman (#31): deepseek/deepseek-r1
Members:
  - #1: anthropic/claude-opus-4.5
  - #12: google/gemini-2.5-pro
  - #21: google/gemini-2.5-flash

> conclave_quick("What is the best approach for...")
[Uses your custom selection]

> conclave_reset
## Custom Conclave Cleared

Configuration

Edit config.py to customize:

Conclave Tiers

Each tier has unique models (no overlap) for proper price/performance differentiation:

# Premium: 6 frontier models for complex questions (~$0.30-0.50/query)
COUNCIL_PREMIUM = [
    "anthropic/claude-opus-4.5",        # Claude Opus 4.5
    "google/gemini-3-pro-preview",      # Gemini 3 Pro
    "x-ai/grok-4",                      # Grok 4 (full reasoning)
    "openai/gpt-5.1",                   # GPT-5.1 (flagship)
    "deepseek/deepseek-v3.2-speciale",  # DeepSeek V3.2 Speciale
    "moonshotai/kimi-k2-thinking",      # Kimi K2 Thinking (1T MoE)
]

# Standard: 4 balanced models (default) (~$0.10-0.20/query)
COUNCIL_STANDARD = [
    "anthropic/claude-sonnet-4.5",      # Claude Sonnet 4.5
    "google/gemini-2.5-pro",            # Gemini 2.5 Pro
    "openai/o4-mini",                   # OpenAI o4-mini
    "deepseek/deepseek-chat-v3.1",      # DeepSeek Chat V3.1
]

# Budget: 4 cheap/fast models (~$0.02-0.05/query)
COUNCIL_BUDGET = [
    "google/gemini-2.5-flash",          # Gemini 2.5 Flash
    "qwen/qwen3-235b-a22b:free",        # Qwen 3 235B (free tier)
    "openai/gpt-4.1-mini",              # GPT-4.1 Mini
    "moonshotai/kimi-k2:free",          # Kimi K2 (free tier)
]

Chairman Rotation

The chairman pool uses reasoning models only (not chat models) for high-quality synthesis:

CHAIRMAN_ROTATION_ENABLED = True
CHAIRMAN_ROTATION_DAYS = 7  # Rotate weekly

CHAIRMAN_POOL = [
    "deepseek/deepseek-r1",          # DeepSeek R1 reasoning
    "openai/o3-mini",                # OpenAI o3-mini reasoning
    "anthropic/claude-sonnet-4",     # Claude Sonnet 4 (strong reasoning)
    "qwen/qwq-32b",                  # Qwen QWQ reasoning model
]

Consensus Thresholds

CONSENSUS_STRONG_THRESHOLD = 0.75   # 75%+ agreement
CONSENSUS_MODERATE_THRESHOLD = 0.50  # 50-75% agreement
CHAIRMAN_TIEBREAKER_ENABLED = True   # Chairman breaks ties

Eval-Light

A standalone benchmark runner for testing and comparing conclave performance across tiers and over time.

Test Suite Overview

The eval suite includes 16 tasks across 9 categories, designed to test different model capabilities:

Category	Tasks	Difficulty	What It Tests
math	2	Easy-Medium	Arithmetic, word problems, step-by-step reasoning
code	2	Easy-Medium	Bug detection, concept explanation, code examples
reasoning	2	Medium-Hard	Syllogisms, multi-step logic puzzles
analysis	2	Medium	Logical fallacies, tradeoff analysis
summarization	2	Medium	Technical docs, business reports
writing_business	2	Easy-Medium	Professional emails, proposals
writing_creative	2	Easy-Medium	Story openings, original metaphors
creative	1	Easy	Analogies with explanations
factual	1	Easy	Science explanations for general audience

Running Evaluations

# Run all 16 tests at standard tier (default)
python eval.py

# Run at different tiers
python eval.py --tier premium    # 6 frontier models (~$0.30-0.50/query)
python eval.py --tier standard   # 4 balanced models (~$0.10-0.20/query)
python eval.py --tier budget     # 4 cheap/fast models (~$0.02-0.05/query)

# Different modes
python eval.py --mode quick      # Stage 1 only (fastest, cheapest)
python eval.py --mode ranked     # Stage 1 + 2 (adds peer rankings)
python eval.py --mode full       # All 3 stages (default, includes synthesis)

# Filter by category
python eval.py --category math
python eval.py --category code
python eval.py --category reasoning

# Don't save results to disk
python eval.py --no-save

# Combine options
python eval.py --tier premium --mode full --category reasoning

Output Format

Results are saved to evals/eval_<tier>_<mode>_<timestamp>.json with:

metadata: Timestamp, tier, mode, chairman model
summary: Success rate, total time, average time per task
results: Per-task details including:
- Individual model responses
- Peer rankings (for ranked/full modes)
- Chairman synthesis (for full mode)
- Consensus level

Example Output

🏛️  Conclave Eval-Light
   Tier: standard | Mode: full | Tasks: 16
--------------------------------------------------

[1/16] Running: math_arithmetic (math)
   ✓ Completed in 12.34s

[2/16] Running: math_word_problem (math)
   ✓ Completed in 15.67s
...

==================================================
📊 EVAL SUMMARY
==================================================
Tier: standard | Mode: full
Chairman: deepseek/deepseek-r1
Tasks: 16/16 successful
Total time: 287.45s
Avg per task: 17.97s

📋 Results by Task:
  ✓ math_arithmetic (easy) - 12.34s
  ✓ math_word_problem (medium) - 15.67s
  ✓ code_debug (easy) - 11.23s
  ...

💾 Results saved to: evals/eval_standard_full_20251204_143052.json

Comparing Tiers

Run the same eval across all tiers to compare model quality vs cost:

python eval.py --tier budget --category reasoning
python eval.py --tier standard --category reasoning
python eval.py --tier premium --category reasoning

Then compare the JSON outputs to see how different model tiers perform on the same tasks.

Use Cases

Scenario	Recommended Tool	Why
"Review this function"	`conclave_ranked`	See which model catches the most issues
"Redis vs PostgreSQL for sessions?"	`conclave_full`	Important decision, need synthesis
"Ideas for this feature"	`conclave_quick`	Fast diverse brainstorming
"Debug this error"	`conclave_quick`	Quick parallel diagnosis
"Rewrite this paragraph"	`conclave_full` + `chairman_preset="creative"`	Creative synthesis
"Is this architecture sound?"	`conclave_full` + `chairman_preset="code"`	Technical synthesis

Example Tool Output

## Conclave Full Result

**Consensus: ✅ STRONG** (75% agreement)

---

### Chairman's Synthesis

_Chairman: deepseek/deepseek-r1_

[Synthesized answer incorporating best points from all models...]

---

### Model Rankings (lower is better)

1. **claude-sonnet-4.5**: 1.50
2. **o4-mini**: 2.00
3. **gemini-2.5-pro**: 2.75
4. **deepseek-v3.1**: 3.75

_First-place votes:_ claude-sonnet-4.5=3, o4-mini=1

Project Structure

conclave-mcp/
├── server.py      # MCP server entry point (5 tools)
├── conclave.py    # Core 3-stage council logic
├── config.py      # Model tiers, chairman rotation, cost estimates
├── eval.py        # Standalone benchmark runner
└── evals/         # Saved evaluation results

Adding Models

OpenRouter supports 200+ models. Find model IDs at https://openrouter.ai/models

# Add to COUNCIL_* lists in config.py
"x-ai/grok-4"                    # xAI Grok
"meta-llama/llama-4-maverick"    # Meta Llama
"mistralai/mistral-large-2"      # Mistral
"deepseek/deepseek-r1"           # DeepSeek reasoning

Important: Keep each tier's models unique (no overlap) for proper differentiation.

How OpenRouter Works

OpenRouter is a unified API gateway—you don't need separate accounts with OpenAI, Google, Anthropic, etc. One API key, one credit balance, access to all models.

Sign up: https://openrouter.ai
Add credits (prepaid, or enable auto-top-up)
Use your single API key for all models

License

MIT

Attribution

Inspired by Andrej Karpathy's llm-council. The original is a web application for interactively exploring LLM comparisons. This project reimplements the council concept as an MCP server for integration with AI-assisted editors, adding consensus protocol and tiebreaker mechanics.