Jupyter AI Agent Magics brings AI agent capabilities to JupyterLab, enhancing its role as the ultimate research tool for AI research, which has historically lacked robust LLM integration. This project enables seamless interaction with large language models within JupyterLab, streamlining workflows for researchers and data scientists. Drawing inspiration from projects like Jupyter AI, Jupyter MCP Server, and Jupyter AI Agents, it provides practical tools to elevate your notebook experience.
Integrating AI agents directly into Jupyter notebooks provides numerous advantages for data scientists, analysts, and researchers:
- Immediate Assistance: Get code suggestions, debugging help, and explanations without leaving your workflow
- Streamlined Development: Reduce context switching between different applications when seeking help
- Knowledge Augmentation: Access expertise across libraries, algorithms, and best practices on demand
- Educational Support: Learn as you work with explanations and guidance tailored to your code
- Productivity Enhancement: Complete analyses faster with automated code generation and optimization
Jupyter notebooks' architecture of small, discrete code cells creates the perfect environment for AI agent collaboration:
- Granular Context: Each cell provides a focused, manageable unit of code for the AI to analyze
- Iterative Interaction: The cell-by-cell execution model matches AI's turn-based conversation pattern
- Progressive Development: Both humans and AI can build solutions incrementally, one logical step at a time
- Immediate Feedback: Results appear directly below each cell, creating tight feedback loops
- Perfect Scope Size: Cells typically contain just enough code to perform a specific task without overwhelming complexity
Jupyter AI Agent Magics offers:
- Simple to use and understand
%%agentand%%chatmagics - Prebuilt set of tools to modify and run Jupyter cells
- Great usability by implementing LLM response streaming, and user callbacks
- Great extensibility to offer option to easilty introduce new tools for Agents to use and even allow Agents to generate tools on their own
- Great flexibility by integrating with top model providers
- Great hackability, the code is easy to understand and modify
Due to underlying dependency issue datalayer/jupyter-nbmodel-client#46 have to pin jupyter related libraries to older versions. It is a good idea to do this in an isolated environment like conda or venv. Assuming that you have conda installed:
To install in editable/development mode:
# Create new environment
conda create --name py312 python=3.12
conda activate py312
# Install package in dev mode
pip install -e .
# Start jupyter as usual:
jupyter labConfiguration is pretty typical, settings are configured throughout Jupyter restarts. Following model providers are supported:
- Anthropic
- Anthropic via Vertex
- OpenAI
For automatic configuration open configure.ipynb notebook.
Model list is available here: https://ai.google.dev/gemini-api/docs/models API key is configured at https://aistudio.google.com/app/apikey It can be provided as environment variable:
export GOOGLE_API_KEY='your-api-key-here'or explicitly through config manager.
from jupyter_ai_agent_magics import config
config_manager = config.ConfigManager.instance
config_manager.model_provider = config.ModelProvider.Google
config_manager.model_name = "gemini-2.5-pro"
config_manager.api_key = "..."Model list is available here: https://docs.anthropic.com/en/docs/about-claude/models/overview API key is configured at https://docs.anthropic.com/en/docs/get-started It can be provided as environment variable:
export ANTHROPIC_API_KEY='your-api-key-here'or explicitly through config manager.
from jupyter_ai_agent_magics import config
config_manager = config.ConfigManager.instance
config_manager.model_provider = config.ModelProvider.Anthropic
config_manager.model_name = "claude-sonnet-4-20250514"
config_manager.api_key = "..."Model list is available here: https://platform.openai.com/docs/pricing API key is configured at https://platform.openai.com/api-keys It can be provided as environment variable:
export OPENAI_API_KEY='your-api-key-here'or explicitly through config manager. Note that recent GPT-5 models might requre organization verification.
from jupyter_ai_agent_magics import config
config_manager = config.ConfigManager.instance
config_manager.model_provider = config.ModelProvider.OpenAI
config_manager.model_name = "gpt-4"
config_manager.api_key = ""Setup is described in https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/claude You'll need working Google Cloud Project, where you explicitly enable specific model.
To connect your project to your VM/devbox run:
gcloud auth application-default loginModel list is available here: https://docs.anthropic.com/en/docs/about-claude/models/overview You need to specify your project and the region where model is available. See this notebook to for the list of regions: https://colab.research.google.com/github/GoogleCloudPlatform/vertex-ai-samples/blob/main/notebooks/official/generative_ai/anthropic_claude_3_intro.ipynb#scrollTo=Y8X70FTSbx7U
from jupyter_ai_agent_magics import config
config_manager = config.ConfigManager.instance
config_manager.model_provider = config.ModelProvider.AnthropicVertex
config_manager.model_name = "claude-sonnet-4@20250514"
config_manager.project_id = "<your project>"
config_manager.location = "us-east5"Register magic as:
from jupyter_ai_agent_magics import magics
magics.load_ipython_extension(get_ipython())See autoregister_agent_magic.ipynb notebook if you'd like to pre-register magic to be auto loaded at Jupyter startup.
Call into LLM and allow it to execute some actions on your behalf using %%agent magic:
%%agent
Write a notebook demonstrating how to train MNIST model.
If you just want to ask LLM some question, and you don't want it to call into any tools, use %%chat magic:
%%chat
Tell my about asyncio in Python.