Llm Council

What elhamid/llm-council MCP server does

The repository implements a local, browser-based council for comparing responses from multiple language models. A submitted prompt is sent to the models listed in the backend configuration through OpenRouter. The application displays the individual answers, collects peer evaluations, and then asks a selected Chairman model to produce the final response.

The elhamid/llm-council MCP server fork adds persistence and auditability changes around this workflow. Completed runs retain Stage 1 answers, Stage 2 rankings, the mapping between anonymous labels and models, aggregate ranking data, and model roles. Conversation history is loaded from the current backend data layout and can also recognize legacy layouts.

The README identifies the implemented role set as Builder, Reviewer, Synthesizer, and Contrarian, with provider-default mappings in the backend. These roles are intended to give participating models different review perspectives before synthesis.

How it works

Processing takes place in three stages:

1. Each configured council model independently answers the user’s prompt.
2. Models review anonymized versions of the other responses and rank them for qualities such as accuracy and insight. Stage 2 removes duplicate judge models, and the label-to-model mapping is preserved for later inspection.
3. The Chairman receives the collected material and writes the final answer shown to the user.

The backend exposes a streaming endpoint that reports stage start and completion events using Server-Sent Events. The frontend is a React and Vite application that presents responses, reviews, and conversation history. Data is stored in JSON files rather than an external database.

Setup and configuration

Python dependencies are managed with uv, while the frontend uses npm. The documented setup runs uv sync for the backend and npm install inside the frontend directory. A root .env file must contain OPENROUTER_API_KEY, and the key must have access to the OpenRouter credits needed for the selected models.

Council membership and the Chairman are configured in backend/config.py. The examples use model identifiers from OpenRouter, but the README indicates that this configuration can be customized. The application can be started with the repository’s start.sh script or by running the backend and frontend development processes separately. Once running, the browser interface is available at http://localhost:5173.

Limitations and notes

The elhamid/llm-council MCP server is described in the supplied material as a local web app, not as a documented MCP tool implementation. No MCP transport, tool list, MCP client configuration, or compatibility with Claude Desktop, Cursor, Windsurf, or Cline is specified. Developers looking specifically for an MCP server should verify the repository’s current code before selecting it for that purpose.

The project is presented as an experimental fork and is provided without an ongoing support commitment. Running it requires an OpenRouter API key and associated usage credits. Because the council calls several models and then performs review and synthesis, costs depend on the configured models and the number of submitted prompts.