构建您的第一个 Agentic AI：MCP + Ollama 工具调用完整指南

Source: Dev.to

Introduction

了解如何构建能够使用工具、做出决策并执行操作的 AI 代理——全部在本地机器上运行。

Background

• Ollama – 本地 AI 模型运行时（类似于 AI 版 Docker）。
• Tool calling – 让大语言模型能够识别何时需要函数、选择合适的工具、传递正确的参数并解释结果。
• MCP (Model‑Connector‑Protocol) – 为 LLM 与工具和数据源连接提供的标准化方式（相当于 AI 的 USB‑C）。FastMCP 是一个 Python 库，简化了 MCP 服务器的构建。

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    YOUR COMPUTER                            │
│                                                             │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐ │
│  │   Ollama     │◄──►│   Python     │◄──►│   FastMCP    │ │
│  │ (AI Brain)  │    │   Client     │    │   Server     │ │
│  └──────────────┘    └──────────────┘    └──────────────┘ │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Flow

1. 你提出一个问题。
2. Python 客户端将问题连同可用工具列表发送给 Ollama。
3. Ollama 决定是否需要使用工具。
4. 客户端通过 FastMCP 调用工具。
5. 结果返回给 Ollama。
6. Ollama 生成最终答案。

Ollama Setup

Linux

curl -fsSL https://ollama.ai/install.sh | sh

macOS

brew install ollama

Windows

Download from and run ollama --version.
You should see something like ollama version 0.1.x. Then start the server:

ollama serve

Keep this terminal open.

Pull a model that supports tool calling

ollama pull llama3.2   # ~2 GB download
ollama run llama3.2

Test it:

>>> Hello! Who are you?
I'm Llama 3.2, an AI assistant...

>>> /bye

Model Support Table

Models without tool support

Attempting to use them yields: Error: does not support tools (status code: 400).

Project Setup

mkdir mcp-ollama-tutorial
cd mcp-ollama-tutorial

# Virtual environment
python -m venv myenv
source myenv/bin/activate   # Linux/macOS
# myenv\Scripts\activate      # Windows

pip install fastmcp ollama requests
python -c "import fastmcp, ollama, requests; print('✅ All packages installed!')"

Building the MCP Server

Create mcp_server.py：

# mcp_server.py
from fastmcp import FastMCP

mcp = FastMCP("My First MCP Server")

@mcp.tool()
def add(a: int, b: int) -> int:
    """Add two numbers together"""
    return a + b

@mcp.tool()
def greet(name: str) -> str:
    """Greet someone by name"""
    return f"Hello, {name}! Welcome!"

@mcp.tool()
def multiply(a: float, b: float) -> float:
    """Multiply two numbers"""
    return a * b

@mcp.tool()
def get_time() -> str:
    """Get the current time"""
    from datetime import datetime
    return datetime.now().strftime("%I:%M %p")

if __name__ == "__main__":
    mcp.run(transport="sse", port=8080)

Run the server:

python mcp_server.py

You should see:

INFO:     Started server process
INFO:     Uvicorn running on http://127.0.0.1:8080
✅ Checkpoint: Your tool server is running!

Keep this terminal open.

Building the Client

Create client_ollama.py：

# client_ollama.py
import json
import ollama
from fastmcp import Client as MCPClient
import asyncio

OLLAMA_MODEL = "llama3.2"
MCP_SERVER_URL = "http://127.0.0.1:8080/mcp"

# ------------------------------------------------------------
# Step 1: Discover available tools from MCP server
# ------------------------------------------------------------
async def load_mcp_tools():
    """Connect to MCP server and get list of available tools"""
    async with MCPClient(MCP_SERVER_URL) as mcp:
        return await mcp.get_tools()

# ------------------------------------------------------------
# Step 2: Send a user query to Ollama, letting it decide on tools
# ------------------------------------------------------------
async def ask_ollama(prompt: str, tools):
    """Send prompt + tool schema to Ollama and get a response"""
    # Ollama expects a JSON schema describing the tools
    tool_schema = {"tools": tools}
    response = await ollama.Chat(
        model=OLLAMA_MODEL,
        messages=[{"role": "user", "content": prompt}],
        options={"tool_schema": json.dumps(tool_schema)},
    )
    return response

# ------------------------------------------------------------
# Step 3: Execute any required tool calls and feed results back
# ------------------------------------------------------------
async def run():
    tools = await load_mcp_tools()
    user_prompt = input("You: ")
    response = await ask_ollama(user_prompt, tools)

    # If Ollama requests a tool call, execute it
    if "tool_calls" in response:
        async with MCPClient(MCP_SERVER_URL) as mcp:
            for call in response["tool_calls"]:
                tool_name = call["name"]
                args = call["arguments"]
                result = await mcp.call(tool_name, **args)
                # Send the result back to Ollama (simplified)
                response = await ollama.Chat(
                    model=OLLAMA_MODEL,
                    messages=[
                        {"role": "assistant", "content": response["content"]},
                        {"role": "tool", "name": tool_name, "content": json.dumps(result)},
                    ],
                )
    print("\nAI:", response["content"])

if __name__ == "__main__":
    asyncio.run(run())

Run the client in a new terminal:

python client_ollama.py

You can now interact with the agent:

You: Hey, greet Alice and then calculate 150 + 75
AI: *Thinking... I need to use the greet tool and the add tool*
AI: Hello, Alice! Welcome! The sum of 150 and 75 is 225.

Running the System

1. Terminal 1 – start Ollama: ollama serve
2. Terminal 2 – run the MCP server: python mcp_server.py
3. Terminal 3 – run the client: python client_ollama.py

How It Works (Deep Dive)

• Tool discovery – the client fetches a JSON schema from the MCP server (/tools).
• Prompt augmentation – the schema is sent to Ollama so the model knows which functions are available.
• Decision – Ollama decides whether a tool call is needed based on the user query.
• Execution – the client invokes the appropriate function via FastMCP (HTTP/SSE).
• Result integration – the tool’s output is fed back to Ollama, which composes the final answer.

Customization

Add new tools by defining additional @mcp.tool() functions in mcp_server.py. Example:

@mcp.tool()
def get_weather(city: str) -> str:
    """Return a simple weather description for the given city."""
    # Placeholder implementation
    return f"The weather in {city} is sunny with 22°C."

Restart the server and the client will automatically discover the new tool.

Troubleshooting

Real Project Ideas

Email Assistant

• Goal: Manage emails (list, read, draft, send) via natural language.
• Tools: list_emails(), read_email(id), draft_reply(id, content), send_email(id).
• Tech Stack: FastMCP, imaplib/smtplib, SQLite for local cache.

Personal Knowledge Base

• Goal: Smart note‑taking with tagging, search, and summarization.
• Tools: add_note(title, body), search_notes(query), summarize_note(id).
• Tech Stack: FastMCP, sqlite3, optional embedding model (e.g., sentence‑transformers).

Finance Manager

• Goal: Track expenses, generate budgets, and answer financial queries.
• Tools: add_transaction(date, amount, category), monthly_report(month).
• Tech Stack: FastMCP, pandas, matplotlib for visual reports.

Smart Home Controller

• Goal: Control IoT devices (lights, thermostat, locks) via voice/text.
• Tools: set_light(room, state), set_temperature(value), lock_door(door).
• Tech Stack: FastMCP, MQTT or local REST APIs of smart devices.

Data Analysis Assistant

• Goal: Load CSV/Excel files, run analyses, and produce charts.
• Tools: load_dataset(path), describe_data(), plot(column_x, column_y).
• Tech Stack: FastMCP, pandas, seaborn/matplotlib.

Study Assistant

• Goal: Generate flashcards, quizzes, and spaced‑repetition schedules.
• Tools: create_flashcards(topic), quiz_user(topic), schedule_review(topic).
• Tech Stack: FastMCP, sqlite3, optional LLM for content generation.

Each project includes actual code snippets (as shown above), example usage, and a recommended tech stack. By following the guide you’ll have a functional, locally‑run AI agent capable of understanding natural‑language requests, deciding which tools to invoke, executing them, and returning intelligent responses—all without external API keys or cloud costs.