多模型议事MCP服务器

Panel MCP Server

An MCP (Model Context Protocol) server that provides LLM council workflow tools for multi-model queries, debates, and reviews.

Overview

Panel MCP Server enables AI clients (like Claude) to orchestrate multi-model workflows. The calling LLM uses these tools to query multiple models in parallel, run structured debates, get critiques, and then synthesize the results itself.

Key insight: The MCP server exposes tools that return structured multi-model responses. The calling LLM decides when/how to use these tools and can further process, compare, or synthesize the collected responses using its own reasoning.

Features

• council_query: Query multiple LLMs in parallel, returning all responses for synthesis
• debate: Run structured adversarial debates between two models
• critique: Get one model to critique another's response
• challenge: Adversarial stress-testing of ideas with multiple models
• query_model: Query a single model directly
• list_models: Discover available models by provider
• health_check: Check provider status and connectivity

Main LLM Participation

The calling LLM can optionally include its own position when using panel tools via the proposedThought parameter:

Provider Support

Dual-Mode Provider System

1. OpenRouter mode (recommended): Use openrouter/ prefix for any of 300+ models

   • openrouter/anthropic/claude-sonnet-5
   • openrouter/openai/gpt-5.6-sol
   • openrouter/meta-llama/llama-3.3-70b-instruct

2. Direct mode: Use provider prefix for direct API calls (lower latency, no fee)

   • openai/gpt-5.6-sol - calls OpenAI API directly
   • anthropic/claude-sonnet-5 - calls Anthropic API directly
   • google/gemini-3.1-pro-preview - calls Google API directly
   • azure/gpt-6-astra - calls your Azure OpenAI / Microsoft Foundry resource

A direct-mode model whose key is missing falls back to OpenRouter, and a model with no known prefix goes to OpenRouter. Set PANEL_ROUTING=strict to make both fail instead: a model then reaches OpenRouter only when it starts with openrouter/. azure/ never falls back.

Every model answer reports servedBy, the provider that actually answered: query_model and council_query responses, challenge findings, critique metadata, and each debate round (affirmativeServedBy, negativeServedBy). An error from the provider call reports it too. query_model and council_query also split latencyMs into queueMs (waiting for a PANEL_MAX_CONCURRENT slot) and modelMs (the provider call, including retries). To compare routes, send the same model both ways in one council_query, for example openrouter/anthropic/claude-sonnet-5 and anthropic/claude-sonnet-5.

Default Panel: GPT-5.6 Sol + Claude Sonnet 5 + Gemini 3.1 Pro (preview)

Installation

npm install panel-mcp-server
# or
pnpm add panel-mcp-server

Configuration

Set at least one API key:

# Required (at least one)
export OPENROUTER_API_KEY=sk-or-...

# Optional direct providers (lower latency)
export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
export GOOGLE_GENERATIVE_AI_API_KEY=...
export MISTRAL_API_KEY=...

# Optional configuration
export PANEL_DEFAULT_MODELS=openai/gpt-5.6-sol,anthropic/claude-sonnet-5,google/gemini-3.1-pro-preview
export PANEL_MAX_CONCURRENT=5
export PANEL_REQUEST_TIMEOUT_MS=300000
export PANEL_ROUTING=auto # or strict: OpenRouter only for openrouter/ models, no fallback
export PANEL_KEEPALIVE_MS=5000 # HTTP keepalive frame interval, min 1000; 0 = don't force SSE

Usage

As MCP Server (stdio)

panel-mcp-server --stdio

As MCP Server (HTTP)

panel-mcp-server --http 8080

Claude Desktop Configuration

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "panel": {
      "command": "npx",
      "args": ["panel-mcp-server", "--stdio"],
      "env": {
        "OPENROUTER_API_KEY": "your-key-here"
      }
    }
  }
}

Tools

council_query

Query multiple LLM models in parallel. Returns all responses for synthesis.

{
  prompt: string,
  models?: string[],  // defaults to GPT-4o, Claude Sonnet 4, Gemini 2.5 Pro
  systemPrompt?: string,
  proposedThought?: string,  // share your thought with the council
  compareMode?: boolean      // ask models to explicitly compare to your thought
}

debate

Run a structured adversarial debate between two models.

{
  topic: string,
  affirmativeModel: string,
  negativeModel: string,
  rounds?: number,  // 1-5, default: 2
  proposedThought?: string,  // share your thought on the topic
  leaningSide?: "affirmative" | "negative" | "neutral"  // which side your thought leans
}

critique

Have one model critique a response.

{
  originalPrompt: string,
  response: string,
  criticModel: string,
  aspects?: string[]  // e.g., ["accuracy", "completeness"]
}

challenge

Have multiple models find weaknesses in a proposed thought. Use for adversarial stress-testing.

{
  proposedThought: string,  // the thought/claim to challenge
  context?: string,         // additional context
  challengers?: string[],   // defaults to GPT-4o, Claude Sonnet 4, Gemini 2.5 Pro
  challengeTypes?: ("logical" | "factual" | "completeness" | "edge_cases" | "alternatives")[]
}

Returns structured challenges with severity ratings and reasoning.

query_model

Query a single model directly.

{
  prompt: string,
  model: string,
  systemPrompt?: string
}

list_models

List available models by provider.

{
  provider?: "all" | "openrouter" | "openai" | "anthropic" | "google" | "mistral"
}

health_check

Check provider status.

{
}

Development

# Install dependencies
pnpm install

# Validate (format + lint + test + build)
pnpm validate

# Development mode
pnpm dev

# Run tests
pnpm test

Built With

• SomaMCP - MCP server framework
• Vercel AI SDK - LLM provider integrations
• functype - Functional programming patterns
• ts-builds - TypeScript build tooling

License

MIT