TaskFlow MCP — Build Your First Production MCP Server

A hands-on lab for the AI Bridge cohort. You'll build, test, and deploy a real MCP server that any LLM can use — Claude, GPT, Gemini, or a free local model via Ollama. Total cost: $0.

What you'll learn

• The three MCP primitives: tools (actions), resources (data), prompts (templates)

• How type hints + docstrings become the LLM's instruction manual

• The dev → prod transport story: stdio locally, Streamable HTTP in the cloud

• Testing MCP servers with an in-memory client (no server process needed)

• Deploying to a free public URL on Render

Related MCP server: mcp-task-manager

Prerequisites

• Python 3.10+ (3.12 recommended)

• uv — pip install uv

• Optional (free local LLM): Ollama with a tool-calling model, e.g. ollama pull qwen3:4b

Part A — Run it locally (5 min)

git clone <repo-url> taskflow-mcp
cd taskflow-mcp
uv sync
uv run taskflow          # starts on stdio, waiting for a client

Press Ctrl+C — stdio servers are meant to be launched by a client, which is next.

Part B — Inspect it (the debugging superpower)

npx @modelcontextprotocol/inspector uv run taskflow

Opens a web UI. Click through Tools → create_task → Run and watch your Python function execute via the protocol. Checkpoint: you can create and list a task in the Inspector.

Part C — Connect a FREE local LLM

pip install mcp-client-for-ollama
ollmcp                    # then add server: command = uv, args = run taskflow

Ask: "Create a high priority task called 'finish the MCP lab', then show me my task stats." Checkpoint: the model calls create_task, then reads tasks://stats.

Works identically with Claude Desktop, Cursor, LM Studio — that's the point of MCP.

Part D — Go to production (Streamable HTTP)

TASKFLOW_TRANSPORT=http uv run taskflow
# → serving at http://localhost:8000/mcp

One env var. Same code. Now it's a network service any client can reach by URL.

Part E — Deploy free on Render

1. Push this repo to your GitHub

2. On render.com: New → Blueprint → pick your repo (it reads render.yaml)

3. Deploy → you get https://<your-app>.onrender.com/mcp

4. Add that URL to any MCP client. Your server is live on the internet. 🎉

Note: free instances sleep when idle; first request after a nap takes ~30–60s.

Run the tests

uv run pytest

Project map

src/taskflow/
├── config.py      # env-driven settings (12-factor)
├── db.py          # SQLite persistence
├── tools.py       # @mcp.tool actions        ← start reading here
├── resources.py   # tasks:// read-only data
├── prompts.py     # user-invoked templates
└── server.py      # FastMCP app + transport switch

Stretch goals

• Add a search_tasks tool with keyword matching

• Add a tasks://overdue resource (you'll need a due_date column)

• Add bearer-token auth for the HTTP transport

• Point two different LLMs at your hosted URL at the same time