DocsWhat you can doUse it from your tools
For agentsOne copy of your data, read by every tool.
Notebooks, Python, DuckDB, Snowflake and Spark read the same open tables. Your agent connects each one with only the access it needs.
Before you start: Claude Code (or any other agent)Data already in LakehouseBoxThe tool you want to use
Say which tools you use
Name the tool and what you want to do with it. Your agent picks the connection and the access level.
My sales are already in LakehouseBox (catalog "sales"). I want to analyse them in a Jupyter notebook with pandas on this Mac, and my colleague asks whether she can read them from Snowflake. Set up the notebook part and tell me what Snowflake needs. Follow lakehousebox.com/docs/
Now this Mac needs your approval to connect. Approve the connection in your browser. Read access is enough for pandas analysis, so I'll give the notebook a read-only key of its own for the catalog sales.
Your turn · in the browser
Approve your agent's access
Open the link your agent gives you (or the page it opens in your browser) and click Allow; when the page asks for a code, type the one your agent gives you. The key it gets acts as you; a notebook that only analyses data can have a read-only key of its own (lhbox token create --name notebook --grant sales:read). You can see and revoke every key on your account page at any time.
Your notebook, ready to run
Your agent builds the notebook, runs it once and keeps the key out of it.
Approved.
The notebook works: I ran it end to end and it reads sales.shop.orders (2,938 rows) into pandas.
- It's in ~/sales-notebook/ with its own Python environment. Open it with
jupyter lab sales.ipynb. - It starts with revenue by region (Norte first, €262k) and revenue by month.
- This Mac is connected read-only to sales. The notebook asks for the key when it runs, so the key never appears in the notebook or on screen.
Example · verified connection
Snowflake and others
Many read the same tables in place, without copying. Some need a one-time step on their side, and not every product can reach them yet.
What Snowflake needs (I couldn't test it here):
- A support case with Snowflake, per account, asking them to allow LakehouseBox's storage address. Until then Snowflake refuses the connection. LakehouseBox's own case was approved the next day.
- An admin of your LakehouseBox account fetches the setup script for the catalog (my connection is read-only and not an admin).
- She runs that script in Snowflake, then adds one statement per table. The data isn't copied; new data appears in Snowflake within about 30 seconds.
Each product has its own page under "For developers and agents → Works with": what it can do with LakehouseBox today, what was tested, and its peculiarities. Some say plainly that a product cannot reach LakehouseBox yet (BigQuery) or has not been tried (Databricks). Your agent reads it before it starts.
Why every tool can read the same data
Your tables are Apache Iceberg tables in open Parquet files. Iceberg is an open standard: DuckDB, PyIceberg, pandas, Polars, Spark, Trino, ClickHouse and Snowflake have been checked against LakehouseBox and read the same tables in place, with no export and no copy that goes stale. Each tool can connect with its own key, limited to the catalogs and the access it needs.
Connect DuckDB on this Mac to my sales, read-only.
Write a notebook cell that charts revenue by month.
Get the Snowflake setup script for my colleague.