mcp-demo

A minimal Python MCP server, built in phases so each commit teaches one concept. End state: Claude Code can call local Ollama models (gemma3:4b / 3:12b) as a tool, routing low-stakes work off the API and onto the homelab.

What is MCP, really?

Model Context Protocol is a standardized JSON-RPC 2.0 protocol that lets an LLM client (Claude Code, Claude Desktop, Cursor) discover and invoke tools, fetch resources, and load prompt templates from separate processes called MCP servers. MCP is to LLM tooling what LSP is to IDE language support: one protocol, many interoperable implementations.

The moving pieces

    ┌───────────────────┐  stdio / HTTP   ┌──────────────────┐
    │  MCP Client       │ ◄──JSON-RPC──► │  MCP Server      │
    │  (Claude Code)    │                │  (this repo)     │
    └───────────────────┘                └──────────────────┘
            │                                     │
            │ spawns as child                     │ hits
            │ process (stdio)                     │ localhost:11434
            ▼                                     ▼
     your shell env                        Ollama / gemma3

Client — embedded in the LLM app
Server — any process that speaks MCP
Transport — stdio (client spawns server as child, pipes JSON-RPC) or streamable-http (server is a web service). This repo uses stdio.

Three primitives an MCP server exposes

Primitive	What it is	This repo's use
Tools	Functions the LLM can call	`echo`, `ollama_ask`, `get_weather`
Resources	Read-only blobs the client can fetch	`ollama://models` (list pulled models)
Prompts	Pre-canned prompt templates	(not used — kept minimal)

Phase progression

Each phase is one commit — git log shows the evolution.

✅ scaffold — pyproject, README, .gitignore.
✅ hello-world server — one echo tool + smoke-test client. Proves the full lifecycle: client spawn → handshake → tool discovery → tool call.
✅ ollama_ask tool — async tool that POSTs to localhost:11434/api/generate and returns Gemma's reply.
✅ polish — adds system parameter to ollama_ask and exposes a resource at ollama://models (list of pulled models).
✅ grounding contrast — adds get_weather (hits NWS api.weather.gov). The point isn't the weather — it's that ollama_ask hallucinates current facts while get_weather returns live data. Same MCP primitive (a tool), completely different epistemic status. This is the difference between an LLM guessing and an agent.

Install & run

python3 -m venv .venv
source .venv/bin/activate
pip install -e .
# Phase 2+ only: run the server by hand to smoke-test
python -m mcp_demo.server

The server reads JSON-RPC off stdin and writes to stdout — if you run it directly in a terminal it will just sit there waiting for input. That's expected. The client (Claude Code) is what actually feeds it.

Registering with Claude Code

MCP servers live in ~/.claude.json (not settings.json — that schema rejects the mcpServers key). The supported path is the claude CLI:

claude mcp add mcp-demo /home/booty/mcp-demo/.venv/bin/python -- -m mcp_demo.server

That writes an entry like this into ~/.claude.json:

"mcp-demo": {
  "type": "stdio",
  "command": "/home/booty/mcp-demo/.venv/bin/python",
  "args": ["-m", "mcp_demo.server"]
}

Restart Claude Code. Tools appear as mcp__mcp-demo__<tool_name> and the resource as ollama://models.

MCP-Demo