# model-council-mcp-server [Health: Active]

**Category:** 🗣️ Conversational AI  
**Repository:** https://github.com/ivanantigravity-lgtm/model-council-mcp-server  
**GitHub Stars:** 1  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/model-council-mcp-server

## Description
Three-model MCP server returning Grok, Gemini, and DeepSeek outputs for Claude.

## Tools
Capabilities this server exposes over MCP:

- **tri_model_scan** — 3 модели отвечают на одну задачу
- **tri_model_compare** — сравнение нескольких вариантов
- **tri_model_red_team** — атака на идею, план или оффер
- **council_model_guide** — краткая памятка по сильным и слабым сторонам моделей

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "model-council-mcp-server": {
    "command": "uvx",
    "args": ["model-council-mcp-server@latest"],
    "env": {
      "POLZA_AI_API_KEY": ""
    }
  }
}
```

**Requires environment variables:** `POLZA_AI_API_KEY` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What model-council-mcp-server MCP server does

The model-council-mcp-server MCP server gives Claude access to a small panel of three language models through the Polza AI API. It submits the same task to all three models concurrently, then returns their concise responses so Claude can produce the final synthesis. The server does not provide the final analysis itself; Claude performs that higher-level summary after receiving the model outputs.

Two presets select different model groups:

- `china`: Moonshot Kimi, Qwen, and DeepSeek
- `usa`: Gemini, Grok, and OpenAI

The configured model IDs can be changed with environment variables if the names in the Polza catalog change.

## How it works

Each tool accepts a `preset` value of `china` or `usa`. For a selected preset, the server sends the task to its three associated models in parallel.

The server prompts the models to stay concise, avoid repeating the question, and omit unnecessary wording. Claude can then compare the independent answers and create an overall response. This approach is intended for ambiguous decisions, evaluation of alternatives, identifying objections, and challenging an idea or plan. A single model is generally more appropriate for simple facts or quick summarization.

## Tools and capabilities

Available MCP tools include:

- `tri_model_scan`: asks three models to respond to the same task.
- `tri_model_compare`: evaluates multiple options or variants.
- `tri_model_red_team`: looks for weaknesses in an idea, plan, or offer.
- `council_model_guide`: provides a short reference to the selected models’ strengths and weaknesses.

The model-council-mcp-server MCP server exposes the same two presets through each tool. The exact model used by each preset is configurable rather than fixed in application code.

## Setup and configuration

The server requires Python 3.11 or newer, `uv`, a Claude Desktop or Claude Code installation, and a `POLZA_AI_API_KEY` value from the Polza API dashboard. A typical MCP configuration runs the package with `uvx` and passes the API key through the server’s environment.

Claude Code uses a `.mcp.json` file in the project root. Claude Desktop uses its platform-specific configuration file: `claude_desktop_config.json` under the macOS application-support directory or the Windows application-data directory. After adding the entry, restart the client and request a council scan.

For local development, clone the repository, run `uv sync`, and start the module with `uv run python -m model_council_mcp_server.server` while providing the API key.

Required configuration:

- `POLZA_AI_API_KEY`

Optional settings include `POLZA_BASE_URL`, model-specific variables for both presets, and `LOG_LEVEL`. The default API endpoint is `https://polza.ai/api/v1`. The Polza model catalog can be queried to confirm current model IDs before overriding them.

## Limitations and notes

The server depends on Polza AI access and a valid API key. Its output consists of short, unaggregated model responses; Claude is expected to perform the final synthesis. Running a council produces three model calls for each request, so it is better suited to questions where independent perspectives matter than to routine lookups or minimal-latency tasks.

The model-council-mcp-server MCP server is distributed under the MIT license.

_Full upstream README: https://allmcps.com/mcp/model-council-mcp-server/readme_