Ollama MCP Server

A local MCP server with a ChatGPT-style web interface for choosing and chatting with Ollama models.

What is included

• server/mcp_server.py — MCP stdio server with 12 tools, three prompts, and an ollama://models resource.

• server/web_tools.py — keyless public web search and readable webpage extraction.

• server/api.py — small HTTP bridge used by the browser, including NDJSON chat streaming.

• frontend/ — responsive React/Vite ChatGPT-style interface.

• setup.ps1, run-ollama.ps1, run-backend.ps1, run-frontend.ps1 — PowerShell launchers.

Related MCP server: Local-MCP-server

Quick start

Open three PowerShell terminals:

cd C:\Users\admin\Desktop\MCP_SERVER
.\setup.ps1

If Ollama is not already running, start it in terminal 1:

cd C:\Users\admin\Desktop\MCP_SERVER
.\run-ollama.ps1

Pull at least one model, for example:

ollama pull llama3.1

Then start the backend and frontend:

# terminal 2
cd C:\Users\admin\Desktop\MCP_SERVER
.\run-backend.ps1

# terminal 3
cd C:\Users\admin\Desktop\MCP_SERVER
.\run-frontend.ps1

Open http://127.0.0.1:5173. The model selector is populated from Ollama, so every installed model can be chosen without changing code.

MCP client configuration

For an MCP client that supports stdio servers, use:

{
  "mcpServers": {
    "ollama-local": {
      "command": "C:\\Users\\admin\\Desktop\\MCP_SERVER\\.venv\\Scripts\\python.exe",
      "args": ["-m", "server.mcp_server"],
      "cwd": "C:\\Users\\admin\\Desktop\\MCP_SERVER"
    }
  }
}

Available MCP tools are list_ollama_models, chat_with_ollama, generate_with_ollama, embed_text, inspect_ollama_model, web_search, scrape_url, calculate, convert, current_time, text_stats, and pretty_print_json.

Available user-invoked MCP prompts are research_topic, solve_math_problem, and summarize_webpage. The server also exposes the read-only ollama://models resource with JSON metadata.

web_search uses DuckDuckGo's HTML endpoint and scrape_url fetches public HTML pages. Both require internet access; local/private network addresses are blocked by the scraper.

The general-purpose tools are local and deterministic: calculate supports safe arithmetic, convert supports length/mass/volume/temperature, current_time supports IANA timezones, text_stats counts text structure, and pretty_print_json validates and formats JSON.

The browser chat binds the 10 safe helper tools from server/tool_registry.py to Ollama's tool-calling API. When a compatible model requests a calculation, web search, scrape, conversion, or other helper, the backend executes it and sends the result back to the model before returning the final answer. The two model-to-model helpers (chat_with_ollama and generate_with_ollama) remain MCP-only to prevent recursive calls.