Connect DuckDB to Excel, Sheets and AI

An analytics engine that happens to be a file

DuckDB is columnar and vectorised, which is a technical way of saying it aggregates millions of rows on a laptop without complaint. That makes it a strange and useful thing to put behind a spreadsheet.

• No server, no service, no port. The file is the database, and the engine runs inside the agent that reads it.
• The aggregation happens in DuckDB, on the machine holding the file. What crosses the network is the result — usually a few hundred rows out of a few hundred million.
• It is the natural end of a data pipeline: someone builds a .duckdb from Parquet or CSV, and this is how the rest of the company reads it without a warehouse being involved.
• Views and tables are both listed, so a file whose useful surface is a set of curated views is queryable as that, not as its raw tables.

Your DuckDB SQL, passed through untouched

There is no translation layer on the agent. The statement you save is the statement DuckDB receives, so the dialect's good parts are available rather than approximated.

• GROUP BY ALL and ORDER BY ALL, which remove most of the tedium from an analytical query.
• SELECT * EXCLUDE (col) and SELECT * REPLACE (...), for when you want everything except the one column nobody needs.
• QUALIFY, so a window function can be filtered without wrapping the whole thing in a subquery.
• Full window functions, CTEs and list comprehensions — DuckDB's dialect, not a portable subset of it.

-- DuckDB SQL, reaching the engine exactly as written
SELECT   * EXCLUDE (internal_id)
FROM (
  SELECT   region, product, sum(amount) AS revenue, internal_id,
           row_number() OVER (PARTITION BY region ORDER BY sum(amount) DESC) AS rk
  FROM     sales
  GROUP BY ALL
)
QUALIFY  rk <= 3
ORDER BY region, revenue DESC

Nested columns arrive as JSON

DuckDB has a richer type system than a spreadsheet cell does, so it is worth knowing what lands where before you build a report on it.

• STRUCT, LIST, MAP and UNION columns come back as JSON text — readable, parseable, and not silently truncated to something wrong.
• BLOB columns are base64 encoded rather than dropped.
• Dates, timestamps, timestamps with time zone and DECIMAL arrive as themselves, so numbers stay numbers and rounding stays yours.
• Anything the driver genuinely cannot represent is labelled as such in the cell instead of arriving as a blank you would have to notice.