// the find
Kaelio/ktx
ktx is an executable context layer for data and analytics agents 🐙 Allow Claude Code, Codex, or other AI agents to query data accurately and with full context of your company
ktx is a local CLI that builds a queryable context layer over your data warehouse — ingesting dbt models, Looker/Metabase definitions, Notion wikis, and raw table introspection into a unified semantic layer with MCP support. It sits between your warehouse and an AI agent so the agent gets approved metric definitions and join-graph awareness instead of reinventing SQL on each prompt. Aimed at analytics engineers and data teams who want Claude Code or Codex to produce consistent, on-policy queries.
Fan/chasm trap detection in the join graph is a real problem most teams just get wrong and hope nobody notices — building automatic resolution into the semantic layer is the right place to fix it. The ingestion model (run once, cache everything, serve via MCP) means agent queries don't repeatedly pay the cost of warehouse introspection. Read-only connection enforcement by design is the correct default for anything touching production data. The project ships actual connector implementations for Postgres, BigQuery, ClickHouse, MySQL, SQLite rather than promising them in a roadmap.
The self-improving context framing is doing a lot of heavy lifting — what actually improves is wiki deduplication and contradiction flagging, not autonomous schema learning, so the pitch slightly oversells the autonomy. Mixed TypeScript + Python runtime (pnpm + uv workspace, plus a separate Python daemon) means your install footprint and failure surface are both larger than they look at first glance. No mention of incremental ingestion cost control: if your dbt project has 500 models, the first `ktx ingest` could burn through a lot of LLM tokens with no obvious budget knob visible in the README. Telemetry is opt-in-by-default with a redaction list that's long enough to suggest the boundaries of what it does collect.