Demo MCP Client Setup

Mainly following the official tutorial from Anthropic: Build an MCP client

⏩ To simply run this demo without going through the tutorial

# First clone repo for MCP Server
git clone https://github.com/MMaghnie/mcp-server

# Then clone this MCP Client repo & navigate to it
git clone https://github.com/MMaghnie/mcp-client.git
cd mcp-client

# Create virtual env
uv venv 

# Activate virtual env (on Windows)
.venv\Scripts\activate

# Install dependencies
uv sync 

# Run the MCP Client and the MCP Server
uv run client.py ..\mcp-server\mcp-server.py 

You can then ask the LLM about the weather.

✅ An example of how a successful run looks like

Here, as an example, the user asked Are there weather alarms for LA?:

(mcp-client) D:\Repos\Playground\MCP\mcp-client>uv run client.py ..\mcp-server\mcp-server.py

Connected to server with tools: ['get_alerts', 'get_forecast']

MCP Client Started!
Type your queries or 'quit' to exit.

Query: Are there weather alarms for LA?

[Calling tool get_alerts with args {'state': 'LA'}]
Yes, there are weather alerts for Louisiana (LA). There are currently **Heat Advisories** in effect:

**Alert 1 - Heat Advisory (Moderate Severity)**
- **Areas affected:** Parts of southeast Arkansas, northeast Louisiana, and central/western Mississippi (including multiple parishes)
- **When:** From 11 AM this morning to 7 PM CDT this evening
- **Heat Index:** Up to 110°F expected
- **Impact:** Hot temperatures and high humidity may cause heat-related illnesses
- **Recommendations:** Stay hydrated, remain in air conditioning, avoid sun exposure, wear lightweight clothing, limit strenuous activities to early morning/evening

**Alert 2 - Heat Advisory (Moderate Severity)**
- **Areas affected:** South central and southwest Arkansas, north central and northwest Louisiana, southeast Oklahoma, and east/northeast Texas
- **When:** From 10 AM to 7 PM CDT Tuesday
- **Heat Index:** Up to 109°F expected
- **Impact:** Hot temperatures and high humidity may cause heat illnesses
- **Recommendations:** Similar precautions as above

Both alerts emphasize the importance of staying cool, hydrating, and checking on relatives and neighbors during this dangerous heat period.

Query: 

ℹ️ Implementation notes for following the tutorial:

These notes are specifically about the Imports and Setup step.

Don't use the most expensive model, as suggested by the tutorial

The tutorial suggests using claude-opus-5 as the MODEL.

But for the purposes of this simple demo, claude-haiku-4-5-20251001* uses fewer tokens and performs well enough.

* This is the most economic model at the time of writing this readme. For an updated list of models, check the official webpage from Anthropic: https://platform.claude.com/docs/en/models/overview and copy the API-ID of the relevant model

If syncing to a remote repo, confirm .env is properly gitignored

The tutorial currently suggests using a command to add .env to .gitignore which doesn't actually let .env be ignored.

Confirm the file is correctly ignored to protect your LLM API key.

Good-to-knows:

• The call to Anthropic() (Line 15) must happen after loading env vars (Line 12), so the LLM API key gets properly populated in the LLM Client SDK.

• Where is the MCP server launched exactly? The server is launched by stdio_client (Line 132), using the description built by server_params (Line 18).

Under the hood - usage of JSON-RPC methods

• Line 132, async with Client(...) as client:, is where the initial server/discover method gets called, for the MCP client and server to get to know each other.

• Line 133, await client.list_tools(), sends the actual tools/list request from the MCP client to the MCP server, effectively asking "what tools do you have?".

• Line 71 executes tools/call

⚠️ Heads-up about extending this demo:

The code in this demo is intentionally kept simple for training purposes and it's not production-ready.

Some reasons:

• The server_params function detects server type simply by a naive file extension check. Anything can be renamed to use any file extension postfix, even malicious files.

• No error handling for server file path.

• Nothing checks what's exactly at the server path before running it.

Read more about best practices for scalability with MCP clients here: Client Best Practices

How the process works at a glance

sequenceDiagram
    actor User
    participant Client as MCP Client<br/>(client.py (process_query))
    participant Server as MCP Server
    participant Claude as Claude API<br/>(language model)

    User->>Client: enters query (chat_loop)

    Client->>Server: tools/list ?
    Server-->>Client: available tools + schemas

    Client->>Claude: messages.create(query, tools=available_tools)

    alt Claude wants to use a tool
        Claude-->>Client: response with tool_use block(s)

        loop for each tool_use block
            Client->>Server: tools/call(name, args)
            Server-->>Client: CallToolResult (content, is_error)
        end

        Client->>Claude: messages.create(messages + tool_results)
        Claude-->>Client: final response (text)
    else Claude answers directly
        Claude-->>Client: response (text only)
    end

    Client-->>User: prints final_text