Cursor Multi Model

License: MIT

Cursor MCP server that queries OpenAI (GPT-4o), Google (Gemini), and Anthropic (Claude Sonnet) in parallel and returns one answer per provider. Register it in Cursor Agent mode and call a single tool — query_llm_models — to get responses from up to three LLMs in one step. Includes a multi-model-research skill for a full compare–reconcile–audit workflow.

Use cases

• Compare GPT, Gemini, and Claude on the same prompt without switching tools
• Run multi-model AI research directly inside Cursor Agent mode
• Get diverse LLM perspectives and reconcile them into one decision-ready output

---

What this MCP does

• One tool: query_llm_models(prompt, models, system_prompt?)
• models: Any non-empty list of "openai", "google", "anthropic" (one, two, or all three).
• Returns: A dict mapping each provider name to its reply text.
• Invocation: Called as a tool call inside Cursor Agent mode.

All configuration (model names and API keys) comes from a .env file. You need your own API keys; this MCP does not provide or store them.

---

Prerequisites

• Python 3.10+
• Cursor (with MCP support)
• API keys for the providers you use: OpenAI, Google AI (Gemini), Anthropic. Usage is billed by each provider; this project does not charge anything.

---

Setup

1. Clone and install

cd cursor_multi_model_mcp
python3 -m venv .venv
source .venv/bin/activate   # On Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Configure environment

• Copy .env.example to .env.
• In .env, set the model names and API keys for the providers you use:
  • OpenAI: OPENAI_MODEL, OPENAI_API_KEY
  • Google: GOOGLE_MODEL, GOOGLE_API_KEY (or GEMINI_API_KEY)
  • Anthropic: ANTHROPIC_MODEL, ANTHROPIC_API_KEY
• Do not commit .env. It is listed in .gitignore. Use .env.example as a template only.

3. Register the MCP in Cursor

Add an entry to ~/.cursor/mcp.json (create the file if it does not exist). Use the full path to this repo on your machine.

{
  "mcpServers": {
    "multi-model-llm": {
      "command": "/full/path/to/cursor_multi_model_mcp/.venv/bin/python",
      "args": ["/full/path/to/cursor_multi_model_mcp/server.py"]
    }
  }
}

Restart Cursor or reload MCP so the new server is picked up.

Windows: Use the full path to .venv\Scripts\python.exe as command and the full path to server.py in args. In JSON, escape backslashes (\\) or use forward slashes. Example (replace C:/Users/You/cursor_multi_model_mcp with your repo path):

"multi-model-llm": {
  "command": "C:/Users/You/cursor_multi_model_mcp/.venv/Scripts/python.exe",
  "args": ["C:/Users/You/cursor_multi_model_mcp/server.py"]
}

---

Examples (query_llm_models)

Below, models (list) → result (dict: provider name → reply text).

One model: ["anthropic"] → {"anthropic": "..."}

Two models: ["openai", "google"] → {"openai": "...", "google": "..."}

All three with system_prompt: prompt "Compare REST vs GraphQL for a mobile backend.", models: ["openai", "google", "anthropic"], system_prompt: "Keep each answer under 200 words." → dict with all three keys.

Deduplication: ["google", "gemini"] → one API call; result has a single key "google".

---

Multi-model-research skill

This repo includes a skill that runs a full multi-model research workflow: problem framing, parallel answers from Cursor + 1–3 MCP models, comparison, reconciliation, evidence escalation, and a final audit.

• Skill file: skills/multi-model-research/SKILL.md
• How to use: Copy skills/multi-model-research to ~/.cursor/skills/multi-model-research so Cursor can load it. In Agent mode, invoke the skill and say which MCP model(s) you want (openai, google, anthropic — one, two, or all three). The agent will call query_llm_models and run the rest of the workflow.

Example: "Use the multi-model-research skill. I want to compare OpenAI, Google, and Anthropic. My question: what's the best way to structure a two-week product discovery sprint?"

Tip: If Cursor is already using Claude, pick only openai and/or google from the MCP so you get different vendors.

---

Run the server manually

For debugging (Cursor normally starts the server via mcp.json):

source .venv/bin/activate
python server.py

The server uses stdio; Cursor talks to it via the command and args in mcp.json.

---

FAQ

Q: How do I call OpenAI, Gemini, and Claude from Cursor in one step?
A: Use the query_llm_models tool from this MCP — pass models: ["openai", "google", "anthropic"] and get one response per provider.

Q: Does this work with Cursor Agent mode?
A: Yes. Register the MCP in ~/.cursor/mcp.json and the query_llm_models tool is available as a tool call inside any Cursor Agent session.

Q: Can I use only one or two providers?
A: Yes. Pass any non-empty subset: ["openai"], ["google", "anthropic"], or all three. Only the providers you include are called.

---

License

This project is licensed under the MIT License. See the LICENSE file.