Keenable's SELECT turns web search into SQL

Keenable's MCP server, called SELECT, treats web search like a database query. One call searches over 1,000 live pages, applying an exact WHERE clause at zero LLM cost. Only the rows that survive get a small LLM call, per SELECT's showcase.

For agent builders, that's the gap between one query and running an LLM over a thousand results just to find the ones that matter.

The usual agent pattern works differently. Ten search results come back, and the model reads each one individually, a cost paid whether or not a given page turns out relevant.

The mechanism is SQL. WEB_SEARCH(), SEM_EXTRACT(), and SEM_MATCH() run as operators inside a normal query, with DuckDB doing the final filtering.

SELECT applies a database habit, ordering the cheap filter before the expensive one, somewhere agents hadn't tried it before.

The showcase doesn't say how SEM_MATCH's precision compares to a plain keyword filter. It also doesn't say what happens when a WHERE clause is too loose and most rows survive to the expensive step anyway. Until there's a public benchmark, this is architecture, not a measured result.