Use with coding agents

This page is for you (not for the agent). If you use Cursor, Claude Code, or another coding assistant, you can point it at this cookbook so it copies the snippets on this site instead of inventing Dash, Streamlit, Reflex, or FastAPI code.

The GitHub repository is databricks-solutions/databricks-apps-cookbook. Full flags and client path tables live in that README.

You need git, bash, and Python 3. Everything below installs from this public clone (the MCP virtualenv uses PyPI). You do not need a Databricks VPN. Node.js 20+ is only if you build the docs website.

:::info When not to use this cookbook

You want	Use instead
Charts and KPIs only, no custom app	AI/BI (Lakeview) dashboards
A natural-language app in the Databricks UI	Genie App Builder
A new TypeScript/React Databricks App	Official AppKit via AI Dev Kit (databricks apps init)
React + FastAPI toolkit	apx
The agent to run SQL, jobs, Unity Catalog, Genie against your workspace	Managed MCP / Unity Gateway, plus databricks aitools install

This cookbook is the Python recipe layer: Dash, Streamlit, Reflex, and FastAPI only.

:::

What you install

Two optional pieces, both from a local clone of the cookbook:

1. Skills — short guides so the assistant picks tables, volumes, auth, and so on.
2. Recipe MCP — tools named list_cookbook_recipes and get_cookbook_recipe so the assistant can pull a snippet and its permissions even if your app folder is empty.

The MCP server only reads this repository. It is not Databricks workspace MCP (no SQL warehouse, no Genie). Keep the clone on disk; client config points at it with absolute paths.

git clone https://github.com/databricks-solutions/databricks-apps-cookbook.git
cd databricks-apps-cookbook

# This folder only (same idea as AI Dev Kit "project" scope)
./install.sh
./mcp-server/mcp_install.sh

# Every repo on this machine
./install.sh --global
./mcp-server/mcp_install.sh --global

Default clients match AI Dev Kit (Claude Code, Cursor, Copilot, Codex, Gemini, and others). For Cursor and Claude Code only:

./install.sh --tools claude,cursor
./mcp-server/mcp_install.sh --tools claude,cursor

A new app folder (do not open the cookbook itself if you want a from-scratch test):

COOKBOOK=/path/to/databricks-apps-cookbook
mkdir -p ~/tmp/cookbook-scratch && cd ~/tmp/cookbook-scratch && git init
"$COOKBOOK/install.sh" --target-dir "$PWD"
"$COOKBOOK/mcp-server/mcp_install.sh" --target-dir "$PWD"

Claude Code can install skills as a plugin instead of a copy: in Claude, /plugin marketplace add databricks-solutions/databricks-apps-cookbook then /plugin install databricks-skills@databricks-skills.

Try it in Cursor

1. File → Open Folder on your app (for a from-scratch test: ~/tmp/cookbook-scratch).

2. Settings → MCP: turn on the cookbook server. Restart Cursor if it is missing. Cursor has no global MCP file (same as AI Dev Kit); the installer writes .cursor/mcp.json in the project.

3. Ask something that should use a recipe:

   Build a Streamlit Databricks App that reads a Unity Catalog table. Use Databricks Apps Cookbook recipes only. Do not invent code.

   You should see app.yaml, a Streamlit layout, and table-read code that matches Read Delta table — not a generic Streamlit tutorial.

4. Ask something that should not use this cookbook:

   I want a TypeScript/React Databricks App.

   The assistant should send you to AppKit / AI Dev Kit, not scaffold Dash or Streamlit from here.

Try it in Claude Code

cd ~/tmp/cookbook-scratch && claude

Confirm a cookbook MCP server, then use the same two prompts as Cursor.

Upgrade on the next release

Copied skills do not update by themselves. The recipe MCP serves whatever is in your clone.

1. Subscribe: on GitHub, Watch → Custom → Releases for databricks-apps-cookbook. Notes: Releases.
2. When a release lands, use the same flags as the first install:

cd /path/to/databricks-apps-cookbook
git fetch --tags
git pull
./install.sh
./mcp-server/mcp_install.sh

Claude plugin: /plugin marketplace update databricks-skills then /reload-plugins.

Official Databricks platform skills (AppKit, jobs, …) stay on databricks aitools update. That is a different product.

Recipes vs skills

You do not get a new skill for every snippet on this site.

• A recipe is one page here (code, permissions, resources) plus the matching sample in GitHub.
• A skill is a category guide (tables, authentication, …) or a workflow (build-app). Adding another tables recipe updates the docs; agents see it through MCP after you git pull. A new skill is only needed for a new category.

How to contribute a recipe: CONTRIBUTING.md.