model-council-mcp-server

# Model Council MCP Server `mcp-name: io.github.ivanantigravity-lgtm/model-council-mcp-server` MCP server that runs the same task through **3 models in parallel** and returns their short answers. Claude then makes a final summary. Two presets: - `china` — Moonshot Kimi, Qwen, DeepSeek - `usa` — Gemini, Grok, OpenAI ## When to Call Useful when: - The task is ambiguous - There's a risk of a single model providing a confident but incorrect answer - Weaknesses, objections, and trade-offs are important - Three different perspectives are needed, not just one answer Not recommended for: - Simple facts - Quick summarization - Tasks that can be solved by a single normal model ## Installation Requirements - `Claude Desktop` or `Claude Code` - [`uv`](https://docs.astral.sh/uv/) - Python 3.11+ - `POLZA_AI_API_KEY` — key obtained from [polza.ai/dashboard/api-keys](https://polza.ai/dashboard/api-keys) Install `uv`: ```bash curl -LsSf https://astral.sh/uv/install.sh | sh ``` ## Installation in 2 Minutes (via PyPI + uvx) ### Claude Code / VS Code Create `.mcp.json` in the project root: ```json { "mcpServers": { "model-council": { "command": "uvx", "args": ["model-council-mcp-server@latest"], "env": { "POLZA_AI_API_KEY": "your-polza-api-key-here" } } } } ``` Restart Claude Code. ### Claude Desktop (macOS) File `~/Library/Application Support/Claude/claude_desktop_config.json`: ```json { "mcpServers": { "model-council": { "command": "uvx", "args": ["model-council-mcp-server@latest"], "env": { "POLZA_AI_API_KEY": "your-polza-api-key-here" } } } } ``` ### Claude Desktop (Windows) File: `%APPDATA%\Claude\claude_desktop_config.json`. Contents are identical. ## Verification After restarting Claude, ask: > Run through model council (usa) the task: Should I add a subscription to my product? Claude should call the `tri_model_scan` tool and return three short answers. ## Tools - `tri_model_scan` — Three models answer one task - `tri_model_compare` — Comparison of multiple options - `tri_model_red_team` — Attack on an idea, plan, or offer - `council_model_guide` — Brief guide on strengths and weaknesses of models Each tool has a `preset` parameter (`china` or `usa`). ## Presets ### `china` - `moonshotai/kimi-k2.5` - `qwen/qwen3.6-plus` - `deepseek/deepseek-v3.2` ### `usa` - `google/gemini-3.1-flash-lite-preview` - `x-ai/grok-4.1-fast` - `openai/gpt-5.4-nano` Verify model IDs through `GET https://polza.ai/api/v1/models/catalog`. If a model is renamed on Polza, substitute its ID using the environment variables below. ## Environment Variables | Variable | Required | Default | | --- | --- | --- | | `POLZA_AI_API_KEY` | yes | — | | `POLZA_BASE_URL` | no | `https://polza.ai/api/v1` | | `COUNCIL_CHINA_MOONSHOT_MODEL` | no | `moonshotai/kimi-k2.5` | | `COUNCIL_CHINA_QWEN_MODEL` | no | `qwen/qwen3.6-plus` | | `COUNCIL_CHINA_DEEPSEEK_MODEL` | no | `deepseek/deepseek-v3.2` | | `COUNCIL_USA_GEMINI_MODEL` | no | `google/gemini-3.1-flash-lite-preview` | | `COUNCIL_USA_GROK_MODEL` | no | `x-ai/grok-4.1-fast` | | `COUNCIL_USA_OPENAI_MODEL` | no | `openai/gpt-5.4-nano` | | `LOG_LEVEL` | no | `INFO` | ## Server Response Compact JSON: - Task - Context - Three raw short answers from models - Sources for each model The server forces models to answer briefly, without filler content or repeating the question, to avoid inflating the Claude context. ## Local Development ```bash git clone https://github.com/ivanantigravity-lgtm/model-council-mcp-server.git cd model-council-mcp-server uv sync POLZA_AI_API_KEY=your_key uv run python -m model_council_mcp_server.server ``` ## License MIT