This is not load balancing — the point is disagreement. Three small models with three different system prompts produce three framings of the same question, and the assistant that called them does the synthesis. Everything runs against your own Ollama instance, so the cost is RAM rather than tokens.
An MCP server that fans one question out to multiple Ollama models running on your machine and returns all of their responses together. Each model can carry a different system prompt, so the answers come back from deliberately different angles rather than as three paraphrases.
- Every Ollama model installed on the system listed, with the configured defaults marked — `list-available-models`
- One question sent to several models at once, their responses returned side by side — `query-models`
- A persona per model, set through per-model system prompts in `.env` — the shipped example gives one creative, one supportive and one analytical framing
- The default roster set once in `DEFAULT_MODELS`, or chosen per question in the request
Ollama installed and running, with the models pulled beforehand — the README's example roster is `gemma3:1b`, `llama3.2:1b` and `deepseek-r1:1.5b`. Node.js 16.x or higher. `OLLAMA_API_URL` points at the Ollama endpoint (`http://localhost:11434` in the example), `DEFAULT_MODELS` lists the roster, and `GEMMA_SYSTEM_PROMPT`, `LLAMA_SYSTEM_PROMPT` and `DEEPSEEK_SYSTEM_PROMPT` set the personas. Install is a clone plus `npm install` and `npm run build`, then launching `build/index.js` with node; the package.json names it `multi-model-advisor` at 1.0.0. Running several models at once takes memory — pick smaller models if a manager model has already claimed most of it.
One command — npx -y @smithery/cli install @YuChenSSR/multi-ai-advisor-mcp --client claude