Query data files. Don't read them.
Agents query CSV, Parquet, and JSON through curated DuckDB primitives and get back only the answer — shaped, capped, and token-costed. No file dumping.
01
Why lacon
Built for AI agents that need to understand data without spending tokens reading raw files.
Primitives
Curated, not raw SQL
describe / sample / profile / aggregate / filter — operations an agent can't get syntactically wrong. Raw SQL stays an escape hatch.
Tokens
~tokens on every response
Every result includes its estimated token cost. Agents know what they spend before deciding what to ask next.
Safety
HITL for the escape hatch
query --show-sql previews resolved SQL before execution. Curated primitives need no confirmation — their SQL is fully determined by parameters.
02
Primitives
Nine curated operations cover what agents ask about data. Each maps to a deterministic SQL pattern — no prompt injection surface.
| Command | Use when the agent asks… |
|---|---|
describe |
"what's in this file?", "what columns?", "how many rows?" |
sample |
"show me some rows", "what does the data look like?" |
count |
"how many rows where X?" |
profile |
"tell me about column X", "any nulls?", "what values?" |
distinct |
"what are the unique values of X?" |
aggregate |
"total/avg X by Y", "group by Z" |
filter |
"show rows where X", projected columns |
find-duplicates |
"are there duplicates?" |
query |
anything the primitives above can't express (HITL required) |
03
Install and run
# install
pip install lacon
pip install lacon[tokens] # adds ~tokens to every response
# Claude Code skill
claude skill install https://github.com/andrii-su/lacon/releases/latest/download/lacon.skill
# quick start
lacon describe sales.csv --pretty
lacon aggregate sales.csv --group-by country --metrics revenue:sum --pretty
lacon query sales.csv "SELECT year, SUM(revenue) FROM {file} GROUP BY year" --show-sql --pretty