Keenable SELECT Turns the Web Into a Queryable SQL Database
A new MCP server integrates DuckDB and semantic operators to allow AI agents to filter thousands of web pages using SQL.
Keenable has launched SELECT, an AI agent and Model Context Protocol (MCP) server that enables users to perform large-scale web searches using SQL. The system allows agents to treat the live web as a structured database, significantly expanding the volume of data an AI can analyze in a single operation.
Implemented as an MCP server, the system centers on a primary tool called `select`. It utilizes DuckDB to execute read-only SELECT statements against live web data, integrating a suite of custom semantic operators. These include WEB_SEARCH, WEB_FETCH, SEM_EXTRACT, SEM_EXTRACT_ALL, SEM_MATCH, SEM_SCORE, and SEM_NORM. To produce final outputs, the architecture employs two specialized agents: a research agent dedicated to data collection and a report agent that generates HTML reports within a sandboxed Python session.
The Shift from Sequential Search
Traditional AI web search typically relies on an agent fetching a small handful of links—often around 10—and reading them sequentially. This process is frequently limited in scope and can be token-expensive, as the LLM must process large amounts of irrelevant text to find specific answers. Keenable SELECT shifts this logic by moving filtering and extraction into a SQL-like query structure. This allows the system to ingest a much broader dataset and apply precise filtering before the LLM ever processes the final rows of data.
Reducing the 'Token Tax'
By treating the web as a queryable database, Keenable SELECT increases the scale of data an agent can analyze in a single turn. This approach reduces the so-called "token tax" associated with reading irrelevant pages, as the heavy lifting of data filtering is handled by the SQL engine rather than the LLM. Consequently, the system can generate comprehensive, data-driven research reports that would be prohibitively slow or expensive to produce using standard Retrieval-Augmented Generation (RAG) or traditional agentic search loops.
Future Implications
As AI agents move toward more complex research tasks, the ability to perform structured queries over unstructured web data represents a significant architectural shift. Observers will be watching to see how this SQL-based approach scales across different types of web content and whether the use of MCP servers becomes the standard for integrating specialized data engines like DuckDB into LLM workflows.