Databricks MCP Server

Host Model Context Protocol (MCP) prompts and tools on Databricks Apps, enabling AI assistants like Claude to interact with your Databricks workspace. Inspiration - https://github.com/databricks-solutions/custom-mcp-databricks-app#

What is this?

This template lets you create an MCP server that runs on Databricks Apps. You can:

• Add prompts as simple markdown files in the prompts/ folder
• Create tools as Python functions that leverage Databricks SDK

A bridge between Claude and your Databricks workspace - you define what Claude can see and do, and this server handles the rest.

How it Works

Architecture Overview

┌─────────────┐       MCP Protocol      ┌──────────────────┐        OAuth         ┌─────────────────┐
│   Claude    │ ◄─────────────────────► │  dba-mcp-proxy   │ ◄──────────────────► │ Databricks App  │
│    CLI      │     (stdio/JSON-RPC)    │ (local process)  │    (HTTPS/SSE)      │  (MCP Server)   │
└─────────────┘                         └──────────────────┘                      └─────────────────┘
                                                ▲                                           │
                                                │                                           ▼
                                                └────────── Databricks OAuth ──────► Workspace APIs

Components

1. MCP Server (server/app.py): A FastAPI app with integrated MCP server that:

   • Dynamically loads prompts from prompts/*.md files
   • Exposes Python functions as MCP tools via @mcp_server.tool decorator
   • Handles both HTTP requests and MCP protocol over Server-Sent Events

2. Prompts (prompts/): Simple markdown files where:

   • Filename = prompt name (e.g., check_system.md → check_system prompt)
   • First line with # = description
   • File content = what gets returned to Claude

3. Local Proxy (dba_mcp_proxy/): Authenticates and proxies MCP requests:

   • Handles Databricks OAuth authentication automatically
   • Translates between Claude's stdio protocol and HTTP/SSE
   • Works with both local development and deployed apps

Prerequisites

• Claude CLI
• Subscription to Databricks apps
• python

Local Development

# Clone and setup
git clone <your-repo>
cd <your-repo>
./setup.sh

# Start dev server
./watch.sh

# Set your configuration for local testing
export DATABRICKS_HOST="https://your-workspace.cloud.databricks.com"
export DATABRICKS_APP_URL="http://localhost:8000"  # Local dev server

# Add to Claude for local testing
claude mcp add databricks-mcp-local --scope local -- \
  uvx --from git+ssh://git@github.com/YOUR-ORG/YOUR-REPO.git dba-mcp-proxy \
  --databricks-host $DATABRICKS_HOST \
  --databricks-app-url $DATABRICKS_APP_URL

## Customization Guide

This template uses [FastMCP](https://github.com/jlowin/fastmcp), a framework that makes it easy to build MCP servers. FastMCP provides two main decorators for extending functionality:

- **`@mcp_server.prompt`** - For registering prompts that return text
- **`@mcp_server.tool`** - For registering tools that execute functions

### Adding Prompts

The easiest way is to create a markdown file in the `prompts/` directory:

```markdown
# Get cluster information

List all available clusters in the workspace with their current status

The prompt will be automatically loaded with:

• Name: filename without extension (e.g., get_clusters.md → get_clusters)
• Description: first line after #
• Content: entire file content

Alternatively, you can register prompts as functions in server/app.py:

@mcp_server.prompt(name="dynamic_status", description="Get dynamic system status")
async def get_dynamic_status():
    # This can include dynamic logic, API calls, etc.
    w = get_workspace_client()
    current_user = w.current_user.me()
    return f"Current user: {current_user.display_name}\nWorkspace: {DATABRICKS_HOST}"

We auto-load prompts/ for convenience, but function-based prompts are useful when you need dynamic content.

Adding Tools

Add a function in server/app.py using the @mcp_server.tool decorator:

@mcp_server.tool
def list_clusters(status: str = "RUNNING") -> dict:
    """List Databricks clusters by status."""
    w = get_workspace_client()
    clusters = []
    for cluster in w.clusters.list():
        if cluster.state.name == status:
            clusters.append({
                "id": cluster.cluster_id,
                "name": cluster.cluster_name,
                "state": cluster.state.name
            })
    return {"clusters": clusters}

Tools must:

• Use the @mcp_server.tool decorator
• Have a docstring (becomes the tool description)
• Return JSON-serializable data (dict, list, str, etc.)
• Accept only JSON-serializable parameters

Deployment

# Deploy to Databricks Apps
./deploy.sh

# Check status and get your app URL
./app_status.sh

Your MCP server will be available at https://your-app.databricksapps.com/mcp/

The app_status.sh script will show your deployed app URL, which you'll need for the DATABRICKS_APP_URL environment variable when adding the MCP server to Claude.

Authentication

• Local Development: No authentication required
• Production: OAuth is handled automatically by the proxy using your Databricks CLI credentials

Examples

Using with Claude

Once added, you can interact with your MCP server in Claude:

Human: What prompts are available?

Claude: I can see the following prompts from your Databricks MCP server:
- check_system: Get system information
- list_files: List files in the current directory
- ping_google: Check network connectivity

Sample Tool Usage

Human: Can you execute a SQL query to show databases?

Claude: I'll execute that SQL query for you using the execute_dbsql tool.

[Executes SQL and returns results]

Project Structure

├── server/                    # FastAPI backend with MCP server
│   ├── app.py                # Main application + MCP tools
│   └── routers/              # API endpoints
├── prompts/                  # MCP prompts (markdown files)
│   ├── check_system.md      
│   ├── list_files.md        
│   └── ping_google.md       
├── dba_mcp_proxy/           # MCP proxy for Claude CLI
│   └── mcp_client.py        # OAuth + proxy implementation
├── client/                  # React frontend (optional)
├── scripts/                 # Development tools
└── pyproject.toml          # Python package configuration

Advanced Usage

Environment Variables

Configure in .env.local:

DATABRICKS_HOST=https://your-workspace.cloud.databricks.com
DATABRICKS_TOKEN=your-token  # For local development
DATABRICKS_SQL_WAREHOUSE_ID=your-warehouse-id  # For SQL tools

Creating Complex Tools

Tools can access the full Databricks SDK:

@mcp_server.tool
def create_job(name: str, notebook_path: str, cluster_id: str) -> dict:
    """Create a Databricks job."""
    w = get_workspace_client()
    job = w.jobs.create(
        name=name,
        tasks=[{
            "task_key": "main",
            "notebook_task": {"notebook_path": notebook_path},
            "existing_cluster_id": cluster_id
        }]
    )
    return {"job_id": job.job_id, "run_now_url": f"{DATABRICKS_HOST}/#job/{job.job_id}"}

Testing Your MCP Server

This template includes comprehensive testing tools for validating MCP functionality at multiple levels.

Quick Verification

After adding the MCP server to Claude, verify it's working:

# List available prompts and tools
echo "What MCP prompts are available from databricks-mcp?" | claude

# Test a specific prompt
echo "Use the check_system prompt from databricks-mcp" | claude

Comprehensive Testing Suite

The claude_scripts/ directory contains 6 testing tools for thorough MCP validation:

Command Line Tests

# Test local MCP server (requires ./watch.sh to be running)
./claude_scripts/test_local_mcp_curl.sh      # Direct HTTP/curl tests with session handling
./claude_scripts/test_local_mcp_proxy.sh     # MCP proxy client tests

# Test remote MCP server (requires Databricks auth and deployment)
./claude_scripts/test_remote_mcp_curl.sh     # OAuth + HTTP tests with dynamic URL discovery
./claude_scripts/test_remote_mcp_proxy.sh    # Full end-to-end MCP proxy tests

Interactive Web UI Tests

# Launch MCP Inspector for visual testing (requires ./watch.sh for local)
./claude_scripts/inspect_local_mcp.sh        # Local server web interface
./claude_scripts/inspect_remote_mcp.sh       # Remote server web interface