┌─────────────────────────────────────────────────────────────────┐
│  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                      │
└─────────────────────────────────────────────────────────────────┘

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

# Install dependencies
uv sync

# Optional: Create .env file for running tests locally
# (Not required for MCP usage - API key is passed via client config)
echo "OPENROUTER_API_KEY=sk-or-v1-your-key-here" > .env

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

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

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

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

> 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

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

# 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

🏛️  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

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

## 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

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

# 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