Persistent
analytical data
for agents.

Sign up in the browser, then give the agent a key confined to one warehouse. The agent asks for its connection recipe: three statements for DuckDB, equivalents for PyIceberg and Spark.

$ tablemere agent create --name reporter \
    --grant <warehouse_id>:write
key shown once; hand it to the agent
$ tablemere connect --warehouse lake --engine duckdb
CREATE SECRET … ATTACH 'w-…' AS lake …

Importing a Parquet file is one statement in the agent's DuckDB. From then on it is an Iceberg table: versioned, compacted hourly, readable by the engines above.

CREATE TABLE lake.garden.readings AS
  SELECT * FROM read_parquet('readings.parquet');
-- a million rows, 18 MB: about 2.3 s from a laptop

Analysis runs in the agent's environment. A result worth keeping becomes a table with one more statement, and any later session or authorised agent opens it.

CREATE TABLE lake.garden.daily_soil_moisture AS
  SELECT CAST(event_time AS DATE) day,
         sensor_id, round(avg(metric_value), 1) avg_moisture
  FROM lake.garden.readings
  WHERE metric_name = 'moisture'
  GROUP BY 1, 2;