// the find
databricks-solutions/ai-dev-kit
Databricks Toolkit for Coding Agents provided by Field Engineering
A toolkit from Databricks field engineering that gives AI coding assistants (Claude Code, Cursor, Copilot, etc.) Databricks-specific context: 75+ MCP tools, markdown skills teaching platform patterns, and a web UI builder app. It's for developers who live outside Databricks notebooks but need to build pipelines, jobs, Unity Catalog objects, and ML workflows against a Databricks workspace.
The MCP server covers real operational surface area — compute, SQL warehouses, pipelines, vector search, Genie — not just toy wrappers. The skills/tools split is clean: markdown skills for pattern guidance, MCP tools for executable actions, independently composable. The install script handles multi-tool config generation (Claude, Cursor, Gemini, Windsurf, etc.) in one pass, which saves real setup time. Supply chain hygiene is visible: they caught and pinned around the litellm incident and publish a full NOTICE.txt with package attributions.
The Databricks License (not MIT/Apache) means you can't fork this for your own platform toolkit without legal review — easy to miss given the permissive-looking README. Project-scope install default means if you run claude from a different directory, nothing works and the error is confusing. The builder app requires deploying a Lakebase database just to get the chat UI running, which is significant infra overhead for what amounts to a dev tool. Test coverage exists but is heavily MLflow/trace-eval focused; the MCP tool implementations themselves have no obvious unit tests, so you're trusting field engineers got the API calls right.