Databricks AI Dev Kit: agent skills for Databricks work in your own editor

What the AI Dev Kit actually solves for Databricks developers

Coding agents know Python and SQL well. They know much less about Databricks specifics: how a Spark Declarative Pipeline is structured, which Unity Catalog objects a job touches, how a multi-task DAG is expressed, or which framework a Databricks App should default to. The AI Dev Kit addresses that gap by shipping curated skills, written by Databricks field engineers, that an agent loads as context while you work in your own editor.

The README frames the choice as two paths. Genie Code is the free, first-party option built into every Databricks workspace, where notebooks, jobs and Unity Catalog data are already in scope. The AI Dev Kit is the other path: Databricks expertise inside Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, and others the README lists as Antigravity, Windsurf and OpenCode. The target reader is someone who already has a coding agent configured and wants Databricks patterns available there rather than switching into a workspace browser tab.

The scope is broad rather than deep in any one area. The README lists Spark Declarative Pipelines with streaming tables, CDC, SCD Type 2 and Auto Loader; Databricks Jobs; AI/BI Dashboards; Unity Catalog; Genie Spaces; knowledge assistants; MLflow experiments; model serving; and Databricks Apps. That breadth is the point: one install covers the surfaces a data engineer or analytics engineer touches in a week.

How skills, the MCP server and the Builder App fit together

The repository is not one program. The top-level layout shows three distinct pieces plus installer scripts: databricks-skills/, databricks-mcp-server/, databricks-builder-app/, databricks-tools-core/, hooks/, and the install.sh and install.ps1 entry points, alongside .claude-plugin/ and .claude/ directories that the agent tooling reads.

Skills are the primary mechanism. They are files an agent pulls into context, so the model gets Databricks-specific instructions rather than generic Python advice. The README notes that the databricks-apps-python skill defaults to AppKit, described as TypeScript plus React and the recommended path for most Apps use cases, and falls back to Python frameworks (Dash, Streamlit, Flask, FastAPI, Gradio, Reflex) when a specific one is needed. That default is a design decision worth knowing about before your agent scaffolds an app you did not ask for.

The MCP server is a separate, standalone component that exposes Databricks actions to AI clients. The README is explicit that skills are recommended instead, and that the MCP server is kept as a foundation for cases where a custom MCP server is required. The Builder App is a web-based UI for Databricks development under databricks-builder-app/. Because these are independent, installing skills does not commit you to the MCP server, and the README's upgrade guidance treats them as things to remove separately.

Installing the AI Dev Kit and running a first real task

The README points at a Quick Start section titled install in existing project, and the repository carries install.sh for Unix-like systems and install.ps1 for Windows. The installer adds Databricks skills for your user or for an existing project folder. Run the shell installer from the repository root:

./install.sh

On Windows, the PowerShell equivalent is install.ps1. After it completes, your agent's configuration should reference the Databricks skills, and the README notes that the installer guides you through the process and delegates to the Databricks CLI for the official AI Tools.

The README also states that AI Tools are installed and kept up to date directly through the Databricks CLI:

databricks aitools install

If you have an older setup, the README's upgrade note says to re-run the Quick Start install in this repo to uninstall older skills and the MCP server before installing the official AI Tools. Do that first rather than layering a new install on top.

For a first task, ask your agent for something the skills cover directly, such as a Databricks Job with a multi-task DAG or a Unity Catalog table definition. The expected result is output that uses Databricks conventions rather than generic Spark. If you want the skills inside a workspace instead of an editor, install_genie_code_skills.py handles uploading selected skills for Genie Code, though the README cautions that much of the public Databricks skills content is already included in Genie Code.

The deprecation you need to read before installing

The most important thing in this repository is a notice, not a feature. The README states that the skills are now delivered as part of Databricks AI Tools, an engineering-owned repository built with field engineering, and that the bundled skill files in this repo are deprecated and kept for reference only.

That has concrete consequences. Several skills were renamed or merged in the official install. databricks-bundles became databricks-dabs, databricks-genie became databricks-genie-agents, and databricks-spark-declarative-pipelines became databricks-pipelines. Two pairs were merged: databricks-lakebase-autoscale and databricks-lakebase-provisioned into databricks-lakebase, and databricks-config into databricks-core. If your notes, scripts or agent configuration reference the old names, they point at deprecated content.

The split of responsibilities is also uneven. The README says the MCP server and Builder App remain in this repository, that the Builder App will keep being developed, and that the MCP server will be maintained and updated on a best-effort basis as GitHub issues are filed. Best-effort maintenance is a real constraint, not a formality. If your workflow depends on the MCP server, you are depending on a component the project itself describes as a fallback to skills.

Where the AI Dev Kit is the wrong tool

If you do not use a coding agent, this repository has nothing for you. Skills are context for a model; without an agent that reads them, the files sit unused. The README's own comparison makes the alternative clear: Genie Code is built into every Databricks workspace, requires no install, and already has your notebooks, jobs and Unity Catalog data in scope. For someone who is comfortable working inside Databricks and has not adopted an external agent, Genie Code is the shorter path, and the README recommends skills over the MCP server for the same reason.

There is also a supply chain dimension. The README describes an audit after a disclosed incident affecting litellm versions 1.82.7 through 1.82.8, and states that the litellm dependency was removed for most usage, remaining only in the test directory for skills evaluation and optimization, pinned to a safe version. That is a reasonable response, but it also tells you the project carries a nontrivial dependency tree and that its own maintainers treat dependency monitoring as ongoing work. Teams with strict third-party review processes should read NOTICE.txt for attribution before rolling this out.

Finally, the licence field is NOASSERTION. The repository has LICENSE.md and NOTICE.md files, but the metadata does not resolve to a standard identifier. If your organisation gates on SPDX identifiers, that is a question to settle with the actual licence text rather than the repository metadata.

Maintenance, upgrades and the cost of staying current

The last push to the default branch was on 2026-08-13, and the most recent release listed is v0.2.0 from 2026-07-28, following v0.1.14 and v0.1.13 in July 2026. The repository is not archived. The release cadence through July and August 2026 suggests active work, but the README itself says the bundled skills are deprecated, so release activity here does not mean the skills you install are the current ones.

Upgrade cost is mostly organisational. The README instructs existing users to re-run the Quick Start install to uninstall older skills and the MCP server before installing the official AI Tools. That is a removal step, not an in-place update, and it means any local edits to skill files are lost unless you keep them elsewhere. Renamed skills add a second pass: references to databricks-bundles, databricks-genie, databricks-spark-declarative-pipelines, databricks-config and the two Lakebase skills need to be updated wherever they appear.

On licensing, the README points to NOTICE.txt for full third-party attribution, and the repository includes LICENSE.md and NOTICE.md. Read those files rather than relying on the NOASSERTION value in the repository metadata. This is not legal advice; the point is that the licence is not resolvable from the metadata alone.

Editorial conclusion

Install the AI Dev Kit if you already write Databricks code in Claude Code, Cursor, Codex, Gemini CLI or Copilot and want field-written patterns for bundles, pipelines, Unity Catalog and Apps close at hand. Skip it if you work entirely inside a Databricks workspace, where Genie Code is the first-party option, or if you want a supported dependency surface: the README says the MCP server is maintained on a best-effort basis as issues are filed, and the bundled skill files are kept for reference only. Before adopting, check the rename table against your current setup, since databricks-bundles is now databricks-dabs and databricks-config merged into databricks-core, and read NOTICE.txt for third-party attribution. Then run the Quick Start in this repo to remove older skills and the MCP server before installing the official AI Tools.

Frequently asked questions

Official sources

1. databricks-solutions/ai-dev-kit on GitHub
2. Issues
3. README
4. Releases