marimo
| Version | 2026.146.1203 |
| Repo | superproject |
marimo reactive notebook + MCP server on port 2718 with GPU OSM analytics, Airflow deps, and the marimo-team skills bundle for AI agents
Installs marimo (a reactive notebook server that also runs as an MCP
server via --mcp on the same port 2718) plus its data stack into a
pixi env at ${HOME}/.pixi/envs/default: GPU-accelerated OSM analytics
deps (Polars-GPU/cuDF, geopandas, quackosm), Apache Airflow Python
deps (the airflow layer ships no pixi env of its own), and the
marimo-team/learn curriculum. The marimo-team/skills bundle is
unpacked read-only at /opt/marimo-skills (MARIMO_SKILLS_DIR) for AI
agents driving marimo over MCP, and a baked marimo.toml turns
auto_instantiate on so cells run the moment the notebook URL opens.
Every claim below is observable: the pixi-env binaries, the skills
directory, the config line, and — at deploy scope — the live notebook
UI plus the MCP endpoint on /mcp/server.
Services
Section titled “Services”marimo
Acceptance plan
Section titled “Acceptance plan”This candy’s plan: — the runnable spec charly check executes against a live deployment. check: steps are idempotent probes; run: steps change state.
| Intent | Step |
|---|---|
run |
mkdir=/workspace |
run |
mkdir=/opt/marimo-skills |
run |
download=https://github.com/marimo-team/skills/archive/refs/heads/main.tar.gz |
run |
mkdir=${HOME}/.config/marimo |
run |
write=${HOME}/.config/marimo/marimo.toml |
check |
the marimo notebook+MCP server binary is installed in the pixi env |
check |
the Apache Airflow CLI is installed in the same pixi env (the airflow layer ships none of its own) |
check |
the marimo-team/skills bundle is unpacked read-only under MARIMO_SKILLS_DIR |
check |
the baked marimo.toml enables auto_instantiate so cells run on page open |
agent-check |
a marimo deployment is running with marimo edit --mcp --mcp-allow-remote on port 2718 |
agent-check |
a client opens the published notebook port and an MCP client pings /mcp/server |
agent-check |
the marimo editor UI is served and the MCP endpoint answers protocol requests on the same port |
check |
command=${HOME}/.pixi/envs/default/bin/python -c “import marimo, torch, transformers, polars, scipy, altair, plotly, statsmodels, duckdb, sqlalchemy, sqlglot, cvxpy, clarabel, daft, wigglystuff, drawdata, marimo_learn, simpy” |
check |
command=${HOME}/.pixi/envs/default/bin/python -c “import geopandas, shapely, fiona, pyproj, quackosm, geoparquet_io, pmtiles, h3, mercantile, folium, contextily, mapclassify, fastmcp, cudf_polars, polars_st, geopolars, geoarrow.pyarrow, geoarrow.pandas, lonboard, freestiler” |
check |
addr=127.0.0.1:${HOST_PORT:2718} |
check |
http=http://127.0.0.1:${HOST_PORT:2718}/ |
check |
mcp=ping |
check |
mcp=list-tools |