marimo-serve

name: marimo-serve description: "Serve every marimo notebook in a project directory under one ASGI host with auto-discovery. Use when the user wants to browse multiple marimo apps at once (localhost, LAN, or tailnet) instead of running `marimo run` per file. Handles expose-on-tailnet via `tailscale serve`." user-invocable: true

Runs one uvicorn process that auto-mounts every *.py notebook in a directory under http://host:port/<mount>/<stem>. Add or remove files — the URL list updates without restart. This is marimo's closest equivalent to "JupyterLab for many notebooks."

Default is read-only Run mode. Pass --edit to launch marimo edit DIRECTORY --watch instead (full editor, saves on disk; picks up external .py edits without restart).

Quick usage

# Read-only Run mode (default)
pixi run python ${CLAUDE_SKILL_DIR}/scripts/serve.py [DIRECTORY] \
    [--host 127.0.0.1] [--port 2718] [--mount /<project>] [--include-code]

# Edit mode — launches marimo's built-in multi-session editor
pixi run python ${CLAUDE_SKILL_DIR}/scripts/serve.py [DIRECTORY] --edit \
    [--host 127.0.0.1] [--port 2718]

Defaults:

DIRECTORY = ./notebooks
Bind = 127.0.0.1:2718
Mode = read-only Run mode. Users interact with UI widgets and see outputs but cannot edit cells. marimo.create_asgi_app() only supports Run mode.
Mount = /<project> (the parent directory's name — e.g. running from ~/projects/mirror with ./notebooks → /mirror/<stem>). marimo requires a non-empty prefix; pass --mount to override. Run mode only — edit mode serves at /.
Source code hidden (--include-code to reveal; read-only regardless)

Expose on tailnet

Keep the server on 127.0.0.1 and let Tailscale Serve handle the tailnet + HTTPS. Because marimo requires a non-empty mount prefix (/apps), the upstream URL must include the same prefix or tailscale will strip it and marimo returns 404:

# mounts /<project> on the tailnet AND forwards /<project> upstream (path preserved)
tailscale serve --bg --https=443 --set-path=/<project> http://127.0.0.1:2718/<project>
# browse: https://<machine>.<tailnet>.ts.net/<project>/<notebook_stem>
tailscale serve status   # see active config
tailscale serve --https=443 --set-path=/<project> off   # remove just this path
tailscale serve reset    # tear down all paths

Use --set-path if / is already mapped to another service — tailscale serve supports multiple path prefixes on the same hostname.

Service-worker conflict (separate port)

If an app already mapped at / (e.g. a PWA) registers a service worker with root scope, that SW will intercept requests to /<project>/* on the same origin and return its own cached shell — marimo never gets the request. Fix: put marimo on a different HTTPS port so it's a separate origin (SWs cannot cross origins):

tailscale serve --bg --https=8443 --set-path=/<project> http://127.0.0.1:2718/<project>
# browse: https://<machine>.<tailnet>.ts.net:8443/<project>/<notebook_stem>

tailscale serve --https=443 off removes ALL paths on :443 — use --set-path=/<path> off to remove a single prefix.

Persistence (macOS launchd)

For always-on serving, wrap in ~/Library/LaunchAgents/com.user.marimo-serve.plist pointing at the serve.py command with KeepAlive=true. Tailscale Serve config survives reboots automatically.

Dependencies

Project must have marimo and uvicorn installed. With pixi:

pixi add marimo uvicorn

When to use another approach

Editing, not serving: pass --edit (this wraps marimo edit <dir>) or call marimo edit <dir> directly.
Fully static hosting (no Python): marimo export html-wasm → drop on any static host.
Single notebook, one-off: marimo run notebook.py is simpler.

Implementation notes

Run mode uses marimo.create_asgi_app().with_dynamic_directory() — the officially documented pattern for multi-notebook ASGI serving. Filenames starting with _ are skipped (treat as private/helper modules). create_asgi_app()'s docstring states it "only works for application that are in Run mode" — that's why edit mode delegates to the marimo edit CLI (os.execv) instead.

Edit mode passes --watch by default so that edits made to the .py file from outside the browser (e.g. by an agent using Edit/Write) are picked up by the running session without a manual reload. This pairs with the marimo-pair skill: one agent edits the file on disk, the user (or another agent) runs cells in the browser.