ETL Skills — MCP Toolkit

When an ETL pipeline breaks, the useful context is scattered across job definitions, run output, schemas, and failure logs. An agent can only help debug it if it can read all of that.

But handing an agent a Databricks token is dangerous: an agent that can run arbitrary SQL can damage the warehouse it was asked to debug.

Any MCP client — Claude Code, Codex, Cursor, Pi, or DeepSeek — talks to etl-agent-tools, a read-only stdio server. The tools are fixed, output is capped, and secrets are scrubbed before a response reaches the model. Portable skills sit on top: diagnose a pipeline, detect schema drift, propose a safe repair.

The server reads Databricks with a personal access token or an OAuth machine-to-machine service principal, dbt from run_results and manifest artifacts, and Airflow through its stable REST API. What comes back is a root-cause brief. This complements Databricks managed MCP servers rather than replacing them.

v0.3 shipped on 2026-10-14 with dbt, Airflow, Unity Catalog context, and OAuth service principals. The npm package is still v0.2.0-beta.1, Databricks inspection.

Safety is the design constraint. The server is read-only by default and output is capped. In Pi, the four write actions (query, pipeline, run, repair) each require interactive confirmation, and raw SQL is rejected outright.

A guided setup command installs the skills for the chosen harness, writes the project config, and records environment-variable names without ever asking for or storing a token. Shipped as an npm beta.