Ganymede

A notebook runtime with a Rust core, and the Mac, iPad and iPhone apps built on it. The apps run the server inside the app; the same core is a command-line server and a library to embed.

• Notebooks as programs. A notebook takes parameters and returns named results, so it runs from the command line, CI, HTTP, a ganymede:// link or Shortcut on the Mac, or another notebook (Notebooks as programs).
• An agent and a person in one notebook. ganymede mcp lets Claude Code or another MCP client work in a notebook as it goes, so the notebook it leaves reruns from the top (Agents).
• TypeScript notebooks on cdp-kernel, on Node, Deno, or inside the iPad and iPhone apps.
• Works with Jupyter. Ganymede serves the same REST and websocket API as jupyter_server and launches kernels from installed kernelspecs, so JupyterLab, VS Code and ipykernel work with it unchanged. The server itself needs no Python.

Named for Jupiter's largest moon, and for the cupbearer who served the gods on Olympus.

Documentation: ganymede-server.com (sources in docs/). Start with the apps, or Install for the command line, step by step from a clean machine with no Python or Jupyter. Worked notebooks are in Examples, and Why Ganymede has the background.

Features

• Mac, iPad and iPhone apps with the server in-process (ganymede-ui); the Mac app is released, the iPad and iPhone apps are on TestFlight.
• Parameters and named results: ganymede execute -p NAME=VALUE --results, POST /api/notebooks/run, and a host API through which a notebook runs others it has been granted (Host API).
• ganymede mcp for agents, with ganymede nb publish to turn a notebook into a page (Agents over MCP).
• The jupyter_server REST and websocket API, so JupyterLab and VS Code connect unchanged. ganymede lab serves JupyterLab itself.
• Any installed kernel: cdp-kernel for TypeScript, ipykernel, or any other kernelspec. The host has no language-specific code.
• Output matching jupyter execute, and notebook saves that match Python's nbformat byte for byte, with atomic writes and checkpoints outside the notebook folder.
• One binary with a random token, loopback-only by default; see Security.
• Local-first: no telemetry, notebooks stay on your machine unless you connect a remote gateway, and crash reporting is off unless you turn it on (Privacy).
• ganymede-engine, a library for embedding, which builds for iOS, Android, and wasm (Embedding).

Quick start

curl -fsSL https://ganymede-server.com/install.sh | sh   # ganymede, plus cdp-kernel if Node 22.18+ is there
ganymede lab install                                     # JupyterLab's frontend, from PyPI, without Python
ganymede lab                                             # opens JupyterLab in the browser

Install

Prebuilt binaries for Linux (x86_64, aarch64), macOS (Apple silicon, Intel), and Windows (x86_64) are on the releases page. Linux and macOS are tested end to end; the Windows build is untested and best effort (Known gaps). The script above downloads the latest, checks its SHA-256, installs ganymede into ~/.local/bin, and installs cdp-kernel if Node 22.18 or later is there. Pass options after sh -s --: --version 0.1.0 pins a release, --bin-dir DIR installs elsewhere, --no-kernel skips cdp-kernel. To build from source: cargo install --path crates/ganymede-cli. Install has each step by hand and the platform notes.

Use

ganymede run -k fake 'hello'                     # a built-in kernel, for trying things out
ganymede run -k cdp-node 'const x = 2; x * 21'   # any installed kernelspec
ganymede console -k cdp-node                     # completion, multiline, Ctrl-C interrupts
ganymede execute notebook.ipynb -o out.ipynb     # like `jupyter execute`
ganymede trust notebook.ipynb                    # like `jupyter trust`
ganymede serve --root-dir . --port 8888          # point VS Code or JupyterLab at the URL
ganymede lab                                     # JupyterLab, opened in the browser

Getting started walks through these, and CLI and configuration lists every command and flag.

Status

Early software. The protocol, kernel host, nbformat, executor, contents, API, and server pass their tests, and an end-to-end suite compares Ganymede against jupyter execute and jupyter server on cdp-kernel and ipykernel. The VS Code Jupyter extension's remote-server flow passes, and JupyterLab's own bundle, served at /lab, runs notebooks in a browser. The engine builds for iOS, Android, and wasm. See the roadmap and known gaps.

The server has no Python runtime or Python dependency; Replacing Python says where Python still appears and why, and Switching from jupyter_server what a move from Jupyter's own server keeps and loses. UIs are clients of the API and live outside the core.

Repository

Cargo workspace at this directory. docs/ and site/ are the website, and validation/ holds scripts that drive real JupyterLab and VS Code clients. Ganymede's own UI, and the shared TypeScript protocol package that used to be frontend/ here, live in tlockney/ganymede-ui.

The crates, and how they depend on each other, are on the Architecture page.

Contributing

Work is tracked in GitHub issues: each task and each bug found gets an issue before work starts, and every pull request links the issue it resolves (Closes #N), which CI checks, and adds what it learned to the development log as a new file in docs/notes/, so open PRs don't conflict there. The Development page has the rules and how to build the docs.

Releases go through a pull request like any other change; Releasing has the steps.

License

MIT; see LICENSE. Every dependency linked into the binary is under a permissive license; about.toml lists the accepted ones, CI fails on any other, and each release archive carries THIRD-PARTY-NOTICES generated by cargo about.