Keenable SELECT: an agent that searches the web in SQL

→

SELECT
  SEM_EXTRACT(content, 'researcher'),
  SEM_EXTRACT(content, 'left lab'),
  SEM_EXTRACT(content, 'joined lab'),
  SEM_EXTRACT(content, 'move month')
FROM WEB_SEARCH(8 diverse queries)
WHERE SEM_MATCH(content,
  'named researcher moving
   between frontier labs, 2025+')

→

The system behind the reports

Keenable SELECT is an MCP server with one main tool: select. The tool runs one read-only DuckDB SELECT statement on live web data. The server runs the web and semantic operators outside DuckDB, puts their output back into the row set, and then runs the final SQL in DuckDB.

A traditional web search gives an agent ten links. The agent must then read each page and build the answer from expensive tokens. SELECT moves this work into the query. One call can search more than 1,000 pages, filter them with an exact WHERE clause at no LLM cost, extract fields with one small LLM call per row, and group the rows.

MCP tools

• select takes DuckDB SELECT queries and returns the rows. The server saves every query result as a result set with an id, and a later query can read from that id.
• generate_html_report takes a brief and result set ids. A report model on the server writes an HTML report from the rows and returns a shareable link.

Semantic operators

The operators live inside normal SQL. The server finds them in the parsed statement, runs them, and replaces them with plain columns. Exact SQL filters run first, so only the surviving rows go to the LLM operators.

WEB_SEARCH and WEB_FETCH can also run per row. Their arguments can use row columns, for example WEB_SEARCH(name || ' founding year').

Main agent

Every report in this gallery comes from two agents: a research agent that uses the MCP server to gather the data, and a report agent that runs inside generate_html_report on the server and writes the page.

The research agent is a plain tool loop: an LLM with the select tool. It writes and runs its own queries until it can answer, and streams its tool calls, results, and answer as events. A follow-up question continues the conversation on top of the stored transcript. Every run in this showcase asks for an HTML report, so the agent ends each answer with the report link.

Report agent

A second agent writes each report on the server. It gets the brief, the rows of the result sets, and an authoring guide. It builds the page in a sandboxed Python session that holds the result sets as dataframes, so the data reaches the page without the model retyping it. After each publish, the server renders the draft and returns screenshots and the page's JavaScript error count; the agent fixes the document and publishes again, under a fixed budget. Only the final draft stays live, published as a link.