Why Ganymede

Ganymede started as a way to write TypeScript notebooks and run them on an iPad. It became a notebook runtime with a Rust core: the Mac, iPad and iPhone apps are the main product, and the same core runs as a command, a server, or behind an agent. Its author builds it for their own use first. Others may find it useful, but it is not trying to win users.

Where it started

The first design, ts-on-ipados, was an iPad client in front of kernels on a home server over a tailnet, then a kernel on the iPad itself. Two things followed from it:

• Keep the Jupyter protocols. The messaging protocol, the REST and websocket API, and nbformat are language-neutral and well specified. Keeping them keeps JupyterLab, VS Code, the existing kernels, and every .ipynb file.
• Replace the server. The reference server, jupyter_server, is a Python program, and Python doesn't travel to an iPad app. A Rust core does: as a binary, as a library linked into an app, and to wasm.

cdp-kernel, a TypeScript kernel extracted from Deno's, grew alongside it, and is the kernel Ganymede is built around. The iPad use case has the original plan and what it decided.

What it is for now

Four priorities, in order:

1. The notebook as a program. A notebook takes parameters, returns named results, and runs from the CLI, the server API, app links, Shortcuts and other notebooks. A kernel can read and run notebooks through the host API, under explicit per-notebook grants; a notebook run from another can't reach more than its caller. (Notebooks as programs)
2. The apps as a full runtime. The apps run notebooks in-process with no external server, with device APIs, automation, widgets and the share extension, and can switch to a remote server from the window. (Ganymede's app)
3. A person and an agent in one live notebook. An agent reads, edits and runs the same live document a person has open. (An agent and a person in one notebook)
4. TypeScript notebooks for TypeScript developers, with cdp-kernel as the first-class kernel.

Not now: Python in the apps through Pyodide, MCP Apps, a native kernel for another language, CurveZMQ. The Roadmap has what comes next under each priority.

Principles

• Only the core is Rust. The protocol, kernel host, notebook format, contents, API and server. User interfaces, ganymede-ui and the apps' Swift shell included, are clients of the API.
• The host knows no language. The host crates name no language outside their tests; kernels come from kernelspecs, and parameters and results are JSON and a MIME type.
• The Jupyter API is a compatibility target. JupyterLab, VS Code and Jupyter's tools should work unchanged; Ganymede adds routes beside the standard ones, not in place of them.
• The file is the source of truth. Saves are atomic and match Python's nbformat byte for byte; live rooms merge outside changes in cell by cell.
• Conservative defaults. A random token, loopback only, crash reporting off. Grants, permissions and "always allow" are scoped as narrowly as they can be.

How it is built

Thomas Lockney sets the direction and decides what ships. Most of the code and docs are written by Claude sessions working from GitHub issues: every change is an issue, a pull request with green CI, and an entry in the development log. Each pull request is also reviewed by a separate Claude session that didn't write it, against its issue and the project's rules. Releases go out only on Thomas's word (Releasing).

The development log is the long record: what each step found, what failed, and why a choice was made.

The arc

Release	Date	What it added
0.9.0	2026-11-11	ganymede nb publish: a notebook or journal as one self-contained HTML page
0.8.0	2026-11-11	Parameters and named results (execute -p --results, POST /api/notebooks/run); the host API with per-notebook grants; the apps' own kernel registered with the server
0.7.0	2026-11-06	One owner per notebook room on a machine; journaling evals; the apps on the landing page
0.6.0	2026-11-06	The ganymede-journal plugin for Claude Code; opt-in crash reporting; kernel logs, memory and CPU
0.5.0	2026-11-03	Document rooms and the execute endpoint; ganymede mcp; terminals; language servers for jupyterlab-lsp
0.4.0	2026-10-20	A container image; the embedding library; the guide to Ganymede's app
0.3.0, 0.2.0	2026-10-18	Hardening for daily use: save preconditions and per-path locks, kernel shutdown and signals, auth and REST parity fixes, release provenance, SECURITY.md
0.1.0	2026-10-17	The server: kernel host, notebook format, executor, contents and the Jupyter REST and websocket API

The changelog lists every change.