Official Python Libraries > MCPs

I use Claude Code daily for my work. After watching Teresa Torres on Peter Yang's podcast running her entire life through Claude Code, I wanted to connect my Google Calendar, Gmail, and Notion to it too. But I kept looking at MCPs as the only solution. Since I didn't trust any of them, I just... didn't do it.

Then I saw David Siegel's claude-canvas demo and asked which MCP he was using. His reply: "Just using official python libraries." I was enlightened 😅 and decided to use each product's official Python libraries.

The MCP Problem

MCPs expose only a subset of what APIs can do. Limited schemas, missing capabilities, poor error handling. I spent more time working around MCP limitations than actually getting work done.

The Python Library Approach

So I built ai-agent-tools - a Python CLI that wraps the actual Google and Notion APIs. The key insight: AI agents don't need fancy protocols. They need CLI tools with JSON output.

Here's what it looks like in practice:

# List my calendar events
aitools google calendar list --days 7 --json

# Check my Notion tasks
aitools notion tasks list DATABASE_ID --status "Todo" --json

# Create a new task
aitools notion tasks create DATABASE_ID "Write blog post" --priority High

Every command supports --json for structured output that Claude can parse. No protocol overhead, no schema limitations, just direct API access.

Architecture: Library + Skill

Architecture diagram: ai-agent-tools connects AI agents to external productivity services through Claude Code skills and official Python libraries from Google and Notion — ai-agent-tools connects AI agents to external productivity services through Claude Code skills and official Python libraries

The clever part is the separation of concerns:

The library is generic: No credentials, no personal configuration. Just Python code that wraps APIs.
The skill is personal: A Claude Code skill file contains YOUR database IDs, YOUR timezone, YOUR preferences.

This means you can clone the library, install it, and then create a simple skill file that tells Claude how to use it for your specific setup:

---
name: my-productivity
description: Use when user asks about tasks, calendar, or email
---

# My Productivity Tools

## Notion Config
- Todo Database ID: abc123...
- Schema: Task (title), Status, Priority, topic, due date

## Commands
List tasks: aitools notion tasks list abc123 --json
Create task: aitools notion tasks create abc123 "Title" --status Todo

Now when I say "add a task to buy groceries", Claude knows exactly which database to use and how to format the command.

Get It

The library is open source: github.com/jnuo/ai-agent-tools

Currently it supports:

Google Calendar and Gmail (OAuth)
Notion tasks and pages (API key)
Example Claude Code skill template
Comprehensive docs on setting up credentials

Contributions are welcome. If you want to add support for another API, fork the repo, add your integration, and send a PR.

If you're frustrated with MCP limitations, give this approach a try. Sometimes the simplest solution - a CLI with JSON output - is the best one.