Jupyter AI Demos

This repo contains a set of demos for the new features of Jupyter AI v3 and associated projects.

Jupyter Server Documents

• Jupyter Server Documents (JSD) is a new project that moves all document and kernel state to Jupyter Server using pycrdt
• Lightening fast document updates
• Low memory usage
• All documents are autosaved by default
• Robust handling for out of band changes on text files and notebooks

Chat files

https://github.com/jupyterlab/jupyter-chat

• Chat files are now a core document type for JupyterLab
• These collaborative chat files enable human/human and human/ai collaboration
• You can organize you chat files on the file system
• The left panel offers a leightweight chat experience
• The launcher and main area offer a full document oriented chat experience
• Jupyter chat supports file attachments through drag and drop

Jupyter AI

https://github.com/jupyterlab/jupyter-ai

• In Jupyter AI v3, everything is an AI persona
• Simply @ mention an AI persona to collaborate with it
• Jupyternaut is the default, and the last mentioned persona is sticky for lightweight threads
• Each chat file will find the nearest .jupyter directory and use it for configuration
• Personas work with file attachments
• Personas can call each other (I think we have a bug in this)

Personas as Python packages

https://github.com/jupyterlab/jupyter-ai/blob/main/packages/jupyter-ai/jupyter_ai/personas/base_persona.py

• Developers can build new personas as Python packages
• Any agentic AI framework is supported (LangGraph, CrewAI, Llama Index, Strands, etc.)
• All that is needed is to subclass BasePersona
• Can include a custom avatar for the AI persona
• Once installed, they are automatically enabled using entry points

Local persona development

• Any Jupyter user can build local personas in a single .py file
• .jupyter/personas/<persona_name>.py files will be picked up dynamically
• Easy debugging in the chat documents
• Refresh local personas using /refresh_personas

Tools

https://github.com/jupyter-ai-contrib/jupyter-ai-tools

• Tool calling is a key dimension of AI persona capabilities
• Jupyter will ship with a default set of tools
• Scope is similar to what Claude Code has
• These tools will edit files using the CRDTs for a full collaborative UX
• Still need to work on tool approval UX
• Still work needed to finish tool implementation

MCP

• Jupyter AI v3 will include built in support for MCP servers
• .jupyter/mcp/config.json for MCP server configuration, follows the standard format
• Jupyter AI will automatically start MCP servers and give clients to AI personas (not completed)
• Needed to avoid MCP servers begin duplicated across different personas and chats

Jupyternaut

• Jupyternaut will continue to be the default AI persona for Jupyter AI
• The default AI persona can also be configured in a config file
• Jupyternaut will ship with the default tools (not completed)
• Goal is to be on par with Claude Code, et al. but with a Jupyter flavor