Ollama Python

Supercharge your AI assistant with local LLM access, run powerful AI models right on your own computer, no internet required.

Python Poetry MCP License: MIT

A Python MCP server that exposes your local Ollama models as tools for AI assistants like Windsurf, VS Code, Claude Desktop, and more.

📚 Full Documentation

---

What It Does

Connect your local LLMs to any MCP-compatible AI assistant. No cloud APIs needed.

Tool	What it does
ollama_chat	Chat with any local model (multi-turn, tool-calling)
ollama_generate	Generate text completions
ollama_embed	Create vector embeddings
ollama_list	List installed models
ollama_show	Inspect model details
ollama_pull	Download new models
ollama_delete	Remove models
ollama_ps	List running models

Quick Start

Prerequisites: Python 3.10+, Ollama running locally

pip install mcp-ollama-python

Windsurf / VS Code

Add to your MCP config (mcp_config.json):

{
  "mcpServers": {
    "ollama": {
      "command": "py",
      "args": ["-m", "mcp_ollama_python"],
      "disabled": false
    }
  }
}

Restart your editor — done. Your AI assistant can now use local Ollama models.

Try It

Type in your AI assistant's chat:

MCP Tool: ollama / ollama_chat — Use model llama3.1 and explain quantum computing

Key Features

• 🔧 8 MCP tools — Full Ollama SDK access
• 🔄 Hot-swap architecture — Drop a file in tools/, it's auto-discovered
• 🎯 Type-safe — Pydantic models throughout
• 🚀 Lightweight — Minimal dependencies, fast startup
• 🔌 Universal — Works with any MCP client

Documentation

Guide	Description
Installation	Setup and prerequisites
Available Tools	All tools with examples
Configuration	Environment variables, model config
Windsurf Integration	Complete Windsurf setup guide
VS Code Integration	VS Code setup
Architecture	How it works, adding tools
Server Control	Start/stop/manage the server
Interactive Manager	Menu-driven management UI
Development	Contributing, code quality
Security	Threat model, host validation, opt-in execute

Security (v1.0.8)

• ollama_execute is opt-in — set OLLAMA_EXECUTE_ENABLED=1; shell/bash removed
• OLLAMA_HOST validation — loopback by default; set OLLAMA_ALLOW_REMOTE_HOST=1 for LAN/remote Ollama
• No HTTP redirects — reduces SSRF risk toward internal endpoints

Details: docs/SECURITY.md

Changelog

1.0.8

• Security hardening: host validation, execute tool gating, env allowlist, input limits
• Fixed __version__ mismatch with package version
• Added tests/test_security.py and docs/SECURITY.md

License

MIT

---