Jupyter AI v3 and the Agent Client Protocol

▶ Watch (0:03)

Jupyter AI v3 launched at this conference. Diamond-Reivich introduced ACP, the Agent Client Protocol. ACP differs from MCP by bringing external AI agents into the Jupyter environment instead of controlling Jupyter from a CLI. The protocol detects installed CLI tools automatically. Users who pip install the Claude Code Python package get the Claude agent inside Jupyter Lab without extra configuration. ACP gives agents access to files, kernels, git, and terminal.

How ACP Brings AI Agents into the Notebook Environment

▶ Watch (2:03)

Jupyter Lab chat is the new interface in v3. It hosts multiple AI personas in one session. Users can tag Claude, Gemini, Codeex, or custom agents simultaneously. Each agent accesses the notebook’s tool registry: running cells, editing code, terminal access, and kernel control. ACP requires agents to ask permission before using tools. The protocol packages all notebook tools into a shared registry that any connected agent can use. This is the mechanism behind Jupyter AI v3’s multi-agent support.

One-Shot Notebook Generation in Jupyter

▶ Watch (4:47)

Diamond-Reivich demonstrated Claude building an astronomy notebook from scratch. The agent installed required packages, generated code cells, and created a Herzsprung-Russell diagram. Claude had context of which packages were already installed. It asked for permissions before executing terminal commands. The entire notebook was generated in one shot. Diamond-Reivich contrasted this with ChatGPT, which hallucinated a non-existent plot library when asked for the same diagram.

Custom AI Personas for Enterprise and Education

▶ Watch (12:19)

Users build custom AI personas by subclassing the base persona class. They define personality and rules in a system prompt, such as enforcing organization color schemes for graphs. The persona manager registry discovers installed agents through Python entry points. Custom personas can restrict tool access. Educational institutions create tutor personas that help students reason about problems without writing code for them. Diamond-Reivich noted the ACP protocol’s permission system allows granular control over what tools the AI can use.

Q&A

What is the difference between using Claude Code in the CLI versus using it inside Jupyter? Jupyter Lab is used by millions as an IDE, and ACP enables multi-agent support within that environment. ▶ Watch (8:00)

How do you configure this for a platform with 100 users? Enterprise cloud systems can connect via API key, and the team is working with Jupyter Hub for larger deployments. ▶ Watch (9:57)

Are there pre-built custom personas available? No pre-built personas exist yet, but the community can build visualization or documentation personas using the system prompt interface. ▶ Watch (13:29)

Notable Quotes

personas. So you can have claude code, Jake Diamond-Reivich · ▶ Watch (2:27)

context of what packages you have or don’t have. Jake Diamond-Reivich · ▶ Watch (6:48)

define what tools the AI has access to. Jake Diamond-Reivich · ▶ Watch (20:46)

Key Takeaways

  • Jupyter AI v3 with ACP embeds AI agents directly inside the notebook environment.
  • ACP detects installed CLI agents automatically with zero manual configuration.
  • Multiple AI personas can collaborate in a single Jupyter Lab chat session.
  • Custom personas with restricted tool access serve education and compliance needs.
  • The protocol requires agents to ask permission before using notebook tools.