"""
Conclave MCP Server Configuration
Customize your conclave members, chairman rotation, and cost controls here.
"""
import os
from datetime import datetime
from typing import Optional
# =============================================================================
# OPENROUTER CONFIGURATION
# =============================================================================
OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1/chat/completions"
# =============================================================================
# COUNCIL TIERS
# =============================================================================
# Premium council for complex questions (architecture, important decisions)
# 6 frontier reasoning models + 1 chairman = 7 (odd)
# Updated December 4, 2025 with latest frontier models
# NOTE: Each tier has UNIQUE models - no overlap between tiers
COUNCIL_PREMIUM = [
"anthropic/claude-opus-4.5", # Claude Opus 4.5 (Nov 2025, 80.9% SWE-bench)
"google/gemini-3-pro-preview", # Gemini 3 Pro (Nov 2025, #1 LMArena 1501 Elo)
"x-ai/grok-4", # Grok 4 (full reasoning, not fast variant)
"openai/gpt-5.1", # GPT-5.1 (flagship reasoning)
"deepseek/deepseek-v3.2-speciale", # DeepSeek V3.2 Speciale (IMO gold, highest compute)
"moonshotai/kimi-k2-thinking", # Kimi K2 Thinking (1T MoE, 256k context, agentic)
]
# Standard council for typical questions (default)
# 4 models + 1 chairman = 5 (odd)
# Updated December 4, 2025
COUNCIL_STANDARD = [
"anthropic/claude-sonnet-4.5", # Claude Sonnet 4.5 (1M context)
"google/gemini-2.5-pro", # Gemini 2.5 Pro (still excellent)
"openai/o4-mini", # OpenAI o4-mini (best reasoning/cost)
"deepseek/deepseek-chat-v3.1", # DeepSeek Chat V3.1 (agent-era model)
]
# Budget council for simple questions (quick checks, brainstorming)
# 4 cheap/fast models + 1 chairman = 5 (odd)
# Updated December 4, 2025
COUNCIL_BUDGET = [
"google/gemini-2.5-flash", # Gemini 2.5 Flash (fast, cheap)
"qwen/qwen3-235b-a22b:free", # Qwen 3 235B (free tier)
"openai/gpt-4.1-mini", # GPT-4.1 Mini (cheap)
"moonshotai/kimi-k2:free", # Kimi K2 (free tier)
]
# Active council (change this to switch tiers, or use tier= parameter in tools)
COUNCIL_MODELS = COUNCIL_STANDARD
# =============================================================================
# COUNCIL MEMBERS (Legacy - use COUNCIL_TIERS above)
# =============================================================================
# Models that participate in the council (Stage 1 opinions + Stage 2 rankings)
# IMPORTANT: Total council size (members + chairman) must be ODD for tiebreaker votes
# By default, chairman is NOT in council (4 members + 1 separate chairman = 5, odd)
# =============================================================================
# CONSENSUS CONFIGURATION
# =============================================================================
# Consensus thresholds
CONSENSUS_STRONG_THRESHOLD = 0.75 # 75%+ agreement = strong consensus
CONSENSUS_MODERATE_THRESHOLD = 0.50 # 50-75% = moderate consensus
# Below 50% = split/weak consensus
# Chairman tiebreaker settings
CHAIRMAN_TIEBREAKER_ENABLED = True # Chairman casts deciding vote on splits
# =============================================================================
# CHAIRMAN CONFIGURATION
# =============================================================================
# Pool of models eligible for chairmanship (Stage 3 synthesis)
# IMPORTANT: By default, chairman should NOT be in COUNCIL_MODELS to ensure odd total
# These are reasoning/thinking models (not chat models) for quality synthesis
# Updated December 4, 2025 - using reasoning models only
CHAIRMAN_POOL = [
"deepseek/deepseek-r1", # DeepSeek R1 reasoning model
"openai/o3-mini", # OpenAI o3-mini reasoning
"anthropic/claude-sonnet-4", # Claude Sonnet 4 (strong reasoning, not in tiers)
"qwen/qwq-32b", # Qwen QWQ reasoning model (32B)
]
# Rotation settings
CHAIRMAN_ROTATION_ENABLED = True
CHAIRMAN_ROTATION_DAYS = 7 # Rotate every N days
# Default chairman when rotation is disabled
DEFAULT_CHAIRMAN = "deepseek/deepseek-r1"
# Title generation model (fast and cheap)
TITLE_MODEL = "google/gemini-2.5-flash"
# =============================================================================
# CONTEXT-BASED CHAIRMAN PRESETS
# =============================================================================
# Use these with chairman_preset="code" etc. in tool calls
# NOTE: These override the default chairman - may result in even council if preset
# model is also in COUNCIL_MODELS. Use explicit chairman= for full control.
# Updated December 4, 2025 - all presets use reasoning models
CHAIRMAN_PRESETS = {
"default": "deepseek/deepseek-r1", # DeepSeek R1 reasoning
"code": "deepseek/deepseek-r1", # Excellent at code reasoning
"creative": "anthropic/claude-sonnet-4", # Creative synthesis with reasoning
"reasoning": "openai/o3-mini", # OpenAI reasoning model
"concise": "qwen/qwq-32b", # QWQ reasoning (efficient)
"balanced": "deepseek/deepseek-r1", # Well-rounded reasoning
}
# =============================================================================
# COST ESTIMATION (approximate USD per 1K tokens)
# =============================================================================
# Input/output costs for estimation (update as pricing changes)
# Updated December 4, 2025
MODEL_COSTS = {
# Format: "model": (input_per_1k, output_per_1k)
# Premium tier (Dec 2025 frontier)
"anthropic/claude-opus-4.5": (0.005, 0.025), # Opus 4.5 (Nov 2025)
"google/gemini-3-pro-preview": (0.00125, 0.01), # Gemini 3 Pro (Nov 2025)
"x-ai/grok-4": (0.003, 0.015), # Grok 4 full (reasoning)
"openai/gpt-5.1": (0.005, 0.015), # GPT 5.1 (flagship)
"deepseek/deepseek-v3.2-speciale": (0.0003, 0.0006), # V3.2 Speciale (high compute)
"moonshotai/kimi-k2-thinking": (0.0006, 0.002), # Kimi K2 Thinking (1T MoE)
# Standard tier
"anthropic/claude-sonnet-4.5": (0.003, 0.015), # Sonnet 4.5
"google/gemini-2.5-pro": (0.00125, 0.01), # Gemini 2.5 Pro
"deepseek/deepseek-v3.1": (0.00014, 0.00028), # DeepSeek V3.1 Terminus
# Budget tier
"x-ai/grok-4.1-fast": (0.0002, 0.0005), # Grok 4.1 Fast (Nov 2025)
"x-ai/grok-4.1-fast:free": (0.0, 0.0), # Grok 4.1 Fast free tier
# Standard tier
"openai/o4-mini": (0.0011, 0.0044), # o4-mini (best reasoning/cost)
"openai/o3-mini": (0.0011, 0.0044), # o3-mini
"deepseek/deepseek-r1-0528": (0.00055, 0.00219),
"deepseek/deepseek-r1": (0.00055, 0.00219),
"google/gemini-2.5-pro": (0.00125, 0.01),
# Budget tier
"google/gemini-2.5-flash": (0.00015, 0.0006),
"deepseek/deepseek-chat-v3-0324:free": (0.0, 0.0), # Free tier
"openai/gpt-4.1-mini": (0.0004, 0.0016),
"qwen/qwen3-32b": (0.0003, 0.0003),
# Chairman pool (reasoning models)
"deepseek/deepseek-r1": (0.00055, 0.00219), # DeepSeek R1 reasoning
"openai/o3-mini": (0.0011, 0.0044), # OpenAI o3-mini reasoning
"anthropic/claude-sonnet-4": (0.003, 0.015), # Claude Sonnet 4
"qwen/qwq-32b": (0.00015, 0.0006), # Qwen QWQ reasoning
}
# Default cost for unknown models
DEFAULT_MODEL_COST = (0.002, 0.01)
# =============================================================================
# COST CONTROLS
# =============================================================================
# Maximum spend per session (set to None to disable)
MAX_SESSION_SPEND_USD: Optional[float] = None # e.g., 1.00 for $1 limit
# Require confirmation for full council queries
REQUIRE_CONFIRMATION_FOR_FULL = False
# =============================================================================
# CHAIRMAN ROTATION LOGIC
# =============================================================================
def get_current_chairman(
override: Optional[str] = None,
preset: Optional[str] = None,
) -> str:
"""
Get the current chairman model.
Priority:
1. Explicit override (full model ID)
2. Preset name (e.g., "code", "creative")
3. Rotating chairman (if enabled)
4. Default chairman
Args:
override: Explicit model ID to use as chairman
preset: Preset name from CHAIRMAN_PRESETS
Returns:
Model ID string for the chairman
"""
# Priority 1: Explicit override
if override:
return override
# Priority 2: Preset selection
if preset and preset in CHAIRMAN_PRESETS:
return CHAIRMAN_PRESETS[preset]
# Priority 3: Rotation
if CHAIRMAN_ROTATION_ENABLED and CHAIRMAN_POOL:
return _get_rotating_chairman()
# Priority 4: Default
return DEFAULT_CHAIRMAN
def _get_rotating_chairman() -> str:
"""Deterministic rotation based on day number."""
# Days since Jan 1, 2025
epoch = datetime(2025, 1, 1)
day_number = (datetime.now() - epoch).days
# Rotation index
rotation_index = (day_number // CHAIRMAN_ROTATION_DAYS) % len(CHAIRMAN_POOL)
return CHAIRMAN_POOL[rotation_index]
def get_rotation_info() -> dict:
"""Get information about current rotation state."""
epoch = datetime(2025, 1, 1)
now = datetime.now()
day_number = (now - epoch).days
days_into_rotation = day_number % CHAIRMAN_ROTATION_DAYS
days_until_next = CHAIRMAN_ROTATION_DAYS - days_into_rotation
return {
"rotation_enabled": CHAIRMAN_ROTATION_ENABLED,
"rotation_period_days": CHAIRMAN_ROTATION_DAYS,
"current_chairman": get_current_chairman(),
"days_until_rotation": days_until_next,
"chairman_pool": CHAIRMAN_POOL,
}
def estimate_cost(
query_tokens: int,
tier: str = "full",
models: list[str] = None,
chairman: str = None,
) -> dict:
"""
Estimate cost for a council query.
Args:
query_tokens: Approximate tokens in the query
tier: "quick" (stage 1), "ranked" (stage 1+2), or "full" (all stages)
models: Council models to use (defaults to COUNCIL_MODELS)
chairman: Chairman model (defaults to current chairman)
Returns:
Dict with cost breakdown and total
"""
models = models or COUNCIL_MODELS
chairman = chairman or get_current_chairman()
def get_cost(model: str) -> tuple[float, float]:
return MODEL_COSTS.get(model, DEFAULT_MODEL_COST)
# Rough estimates for response sizes
avg_response_tokens = 500
ranking_tokens = 300
synthesis_tokens = 800
cost_breakdown = {
"stage1": 0.0,
"stage2": 0.0,
"stage3": 0.0,
"total": 0.0,
}
# Stage 1: Each model gets query, produces response
for model in models:
input_cost, output_cost = get_cost(model)
cost_breakdown["stage1"] += (
(query_tokens / 1000) * input_cost +
(avg_response_tokens / 1000) * output_cost
)
if tier in ["ranked", "full"]:
# Stage 2: Each model evaluates all responses
eval_input_tokens = query_tokens + (avg_response_tokens * len(models))
for model in models:
input_cost, output_cost = get_cost(model)
cost_breakdown["stage2"] += (
(eval_input_tokens / 1000) * input_cost +
(ranking_tokens / 1000) * output_cost
)
if tier == "full":
# Stage 3: Chairman synthesizes everything
synthesis_input = query_tokens + (avg_response_tokens * len(models)) + (ranking_tokens * len(models))
input_cost, output_cost = get_cost(chairman)
cost_breakdown["stage3"] = (
(synthesis_input / 1000) * input_cost +
(synthesis_tokens / 1000) * output_cost
)
cost_breakdown["total"] = sum([
cost_breakdown["stage1"],
cost_breakdown["stage2"],
cost_breakdown["stage3"],
])
return cost_breakdown
# =============================================================================
# COUNCIL VALIDATION
# =============================================================================
def validate_council_size(models: list[str] = None, chairman: str = None) -> dict:
"""
Validate that council size is odd (including chairman).
Args:
models: Council models (defaults to COUNCIL_MODELS)
chairman: Chairman model (defaults to current chairman)
Returns:
{
"valid": bool,
"total_size": int,
"council_members": int,
"chairman_included": bool,
"message": str
}
"""
models = models or COUNCIL_MODELS
chairman = chairman or get_current_chairman()
council_count = len(models)
chairman_in_council = chairman in models
# Total voting members: council + chairman (if not already in council)
total_size = council_count if chairman_in_council else council_count + 1
is_odd = total_size % 2 == 1
if is_odd:
message = f"Council size is valid: {total_size} members (odd)"
else:
message = (
f"WARNING: Council size is {total_size} (even). "
f"Add or remove 1 model for proper tiebreaker support. "
f"Current: {council_count} council members + "
f"{'chairman already in council' if chairman_in_council else '1 chairman'}"
)
return {
"valid": is_odd,
"total_size": total_size,
"council_members": council_count,
"chairman_included": chairman_in_council,
"message": message,
}
def get_recommended_council_size(current_count: int, chairman_in_council: bool) -> dict:
"""
Recommend how to adjust council size to be odd.
Returns options for adding or removing models.
"""
total = current_count if chairman_in_council else current_count + 1
if total % 2 == 1:
return {"adjustment_needed": False, "current_total": total}
return {
"adjustment_needed": True,
"current_total": total,
"options": [
f"Add 1 model to council (total: {total + 1})",
f"Remove 1 model from council (total: {total - 1})" if current_count > 2 else None,
"Include chairman in council models" if not chairman_in_council else None,
],
}
# =============================================================================
# COUNCIL TIER SELECTION
# =============================================================================
def get_council_by_tier(tier: str = "standard") -> list[str]:
"""
Get council models by tier name.
Args:
tier: "premium", "standard", or "budget"
Returns:
List of model IDs for that tier
"""
tiers = {
"premium": COUNCIL_PREMIUM,
"standard": COUNCIL_STANDARD,
"budget": COUNCIL_BUDGET,
}
return tiers.get(tier.lower(), COUNCIL_STANDARD)
def get_tier_info() -> dict:
"""Get information about available council tiers."""
return {
"premium": {
"models": COUNCIL_PREMIUM,
"count": len(COUNCIL_PREMIUM),
"description": "Frontier models for complex questions",
"estimated_cost": "~$0.30-0.50 per full query",
},
"standard": {
"models": COUNCIL_STANDARD,
"count": len(COUNCIL_STANDARD),
"description": "Balanced models for typical questions",
"estimated_cost": "~$0.10-0.20 per full query",
},
"budget": {
"models": COUNCIL_BUDGET,
"count": len(COUNCIL_BUDGET),
"description": "Fast/cheap models for simple questions",
"estimated_cost": "~$0.02-0.05 per full query",
},
}
