kaelio/ktx

ktx is a tool that helps AI coding assistants like Claude Code, Codex, or Cursor query your company's data warehouse accurately. The core problem it solves is simple: when you ask an AI agent to pull data or answer an analytics question, it usually doesn't know your business logic. It might guess how tables connect, invent its own definition of "revenue," or return numbers that don't match what your finance team actually uses. ktx bridges that gap by giving the agent a ready-made understanding of your metrics, table relationships, and business rules before it ever writes a line of SQL. At a high level, it works by scanning your database, your existing analytics tools (like dbt, Looker, or Metabase), and even your team wikis or Notion docs. It pulls all of that together into a local knowledge base, part structured definitions, part searchable wiki. When an agent then needs to answer a question, it goes through ktx rather than guessing. ktx hands it the approved metric definitions, figures out the right table joins, and generates read-only SQL that runs against your warehouse. It also flags contradictions it finds across your sources, so a human can review them. This is aimed at data teams and analytics engineers who want to let AI agents handle data tasks without babysitting every query. If your company's knowledge about metrics is scattered across dbt models, Looker dashboards, and random Notion pages, ktx consolidates that into one place agents can actually use. It's not for one-off queries, it's for teams who want agents to repeatedly and reliably answer questions using the same definitions everyone else in the company uses. A few things stand out. It runs entirely locally, so your schema and query results don't go to any hosted service, only whatever you send to your chosen LLM provider. It's read-only by design, meaning it can't accidentally modify your database. And it doesn't replace tools like dbt's semantic layer, it ingests them and adds context on top, giving agents a single surface to search rather than juggling three disconnected ones.