Jupyter Ecosystem Developments 2025-2026

Snapshot of the notable shifts in the Jupyter ecosystem over roughly mid-2025 to mid-2026. Two through-lines: (a) agentization of notebooks via open protocols, and (b) a server-side re-plumbing for real-time collaboration, performance, and large-output handling. Relevant to JupyterLab Claude Code.

1. Agents in notebooks — Jupyter AI v3

The dominant story. Jupyter AI was fully rearchitected for v3.0 (RC through early 2026, 3.0.0 targeted end of March 2026).

• Personas — define a skill-specific agent in a few lines of Python; both AI agents and humans are @-mentionable in a shared chat. Shift from one Jupyternaut chatbot toward an ecosystem of domain agents. (issue #1313, The New Stack)
• Agent Client Protocol (ACP) — the headline. ACP is to coding agents what LSP is to language servers: an open standard so any agent plugs into any editor. Jupyter AI v3 pivoted to being an ACP client (jupyter-ai-acp-client), enabling Claude Code, Codex, Gemini, Goose, Kiro, OpenCode etc. inside JupyterLab. (issue #1531, TFiR)

A dedicated org, jupyter-ai-contrib, now houses these human+AI collaboration building blocks as modular extensions instead of one monolith.

2. Server-side re-plumbing — jupyter-server-documents

jupyter-server-documents (also jupyter-ai-contrib): a server-side rewrite of document + kernel handling with a new optimized RTC engine, large memory/perf wins, and much better handling of large cell outputs. Ships NextGenKernelManager. It’s the foundation the v3 AI stack sits on.

3. Jupyter + MCP (two distinct things)

• datalayer/jupyter-mcp-server hit v1.0 in April 2026 (v1.0.2 Apr 27). External MCP server that drives notebooks: smart re-execution on cell failure, multimodal (image/plot) outputs, multi-notebook switching. v1.0 breaking change requires MCP_TOKEN in the client config; v1.0.2 bundles pycrdt.
• jupyter_server_mcp — the inverse: a server extension that makes JupyterLab itself expose an MCP server (via FastMCP) with tools like add_cell, run_cell, execute_command. Also gives v3 agents MCP-based tool access.

4. Core JupyterLab releases

• 4.5 / Notebook 7.5 (Oct 2025) — built-in audio and video viewers; 51 features + 81 fixes. (Jupyter Blog)
• 4.6 (June 18, 2026) — (changelog)
  • contentVisibility is now the default windowing mode (fixes cell-rendering bugs).
  • Minimap for all notebooks/windowing modes (View → Show Minimap).
  • Open a notebook without starting a kernel (Open With); streamed uploads make saving large notebooks much faster.
  • Inline completion ghost text with syntax highlighting (Settings → Inline Completer).
  • Safety: 1-6 no longer accidentally convert code cells to Markdown headings.
  • In-app language pack installation.

Personal relevance

The local dotfiles JupyterLab (nix/home-manager/modules/jupyter.nix) already runs the v3 alpha stack: persona manager, ACP client, jupyter_server_documents (source of the NextGenKernelManager FutureWarning), and jupyter_server_mcp. Those startup warnings are a consequence of being on the bleeding edge, not a broken install. LSP is ty (registered via jupyter_server_config.py, spawns lazily on first Python doc open).