AI/agent privacy controls for sensitive variables and env vars

[ytrofimov]

Marimo’s runtime-aware AI is powerful, but I’m concerned about using it in notebooks that import credentials, read env vars, or create authenticated clients.

It would be useful to have a way to mark variables, env vars, outputs, or objects as unavailable to AI assistants / marimo pair.

Is this an active exploration topic?

Possible API:

token = mo.private(os.environ["DATABRICKS_TOKEN"])
conn = mo.private(sql.connect(...))

Or config:

[ai.privacy]
redact_variables = ["*_TOKEN", "*_SECRET", "*_KEY", "*credentials*", "conn"]
redact_env = ["*_TOKEN", "*_SECRET", "*_KEY", "GOOGLE_APPLICATION_CREDENTIALS"]
redact_outputs = true

This is currently the main thing preventing me from using Marimo as a daily driver alongside or instead of my IDE + AI workflow, because many of my notebooks interact with credentialed services.

Main ask: let users benefit from runtime-aware AI without exposing every live variable, credentialed object, or sensitive output to the model/agent.