The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides bash commands to run the marimo server with flags that reduce the security posture of the local environment.
The command marimo edit --host 0.0.0.0 --no-token --sandbox notebook.py binds the server to all network interfaces (0.0.0.0) and disables the authentication token (--no-token).
This allows any device on the local network to access and potentially edit the notebook.
The skill provides a clear warning in the 'Gotchas' section regarding the risks of LAN exposure when using these flags.
[EXTERNAL_DOWNLOADS]: The skill utilizes PEP 723 script metadata to manage notebook-specific dependencies.
Evidence: The # /// script block includes marimo, polars, jupyter-scatter, and pandas.
When executed in --sandbox mode, the environment will automatically download and install these packages using the uv tool.
These are well-known data science packages and no non-standard or suspicious sources are specified.
[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for loading external data files and building UIs that react to user selection, creating an ingestion surface for untrusted data.
Ingestion points: The pl.read_parquet("data.parquet") command in the 'Notebook skeleton' section.
Boundary markers: None are present in the provided templates to distinguish between data and instructions.
Capability inventory: The skill utilizes mo.ui.table and interactive jscatter widgets which process and display the loaded data.
Sanitization: No specific sanitization or validation of the Parquet file content is described.
This represents a standard surface for indirect prompt injection common in data analysis tools.

marimo-notebook