Databricks MCP Server

A production-ready Model Context Protocol (MCP) server that exposes Databricks REST capabilities to MCP-compatible agents and tooling. Version 0.4.4 introduces structured responses, resource caching, retry-aware networking, and end-to-end resilience improvements.

Table of Contents

1. Key Capabilities

2. Architecture Highlights

3. Installation

4. Configuration

5. Running the Server

6. Integrating with MCP Clients

7. Working with Tool Responses

8. Available Tools

9. Development Workflow

10. Testing

11. Publishing Builds

12. Support & Contact

13. License

Key Capabilities

• Structured MCP Responses - Each tool returns a CallToolResult with a human-readable summary in content and machine-readable payloads in structuredContent that conform to the tool’s outputSchema.

• Resource Caching - Large notebook/workspace exports are cached once and returned as resource_link content blocks with URIs such as resource://databricks/exports/{id} (also reflected in metadata for convenience).

• Progress & Metrics - Long-running actions stream MCP progress notifications and track per-tool success/error/timeout/cancel metrics.

• Resilient Networking - Shared HTTP client injects request IDs, enforces timeouts, and retries retryable Databricks responses (408/429/5xx) with exponential backoff.

• Async Runtime - Built on mcp.server.FastMCP with centralized JSON logging and concurrency guards for predictable stdio behaviour.

Architecture Highlights

• databricks_mcp/server/databricks_mcp_server.py - FastMCP server with tool registration, progress handling, metrics, and resource caching.

• databricks_mcp/core/utils.py - HTTP utilities with correlation IDs, retries, and error mapping to DatabricksAPIError.

• databricks_mcp/core/logging_utils.py - JSON logging configuration for stderr/file outputs.

• databricks_mcp/core/models.py - Pydantic models (e.g., ClusterConfig) used by tool schemas.

• Tests under tests/ mock Databricks APIs to validate orchestration, structured responses, and schema metadata without shell scripts.

For an in-depth tour of data flow and design decisions, see ARCHITECTURE.md.

Installation

Prerequisites

• Python 3.10+

• uv for dependency management and publishing

Quick Install (recommended)

Register the server with Cursor using the deeplink below - it resolves to uvx databricks-mcp-server@latest and picks up future updates automatically.

cursor://anysphere.cursor-deeplink/mcp/install?name=databricks-mcp&config=eyJjb21tYW5kIjoidXZ4IiwiYXJncyI6WyJkYXRhYnJpY2tzLW1jcC1zZXJ2ZXIiXSwiZW52Ijp7IkRBVEFCUklDS1NfSE9TVCI6IiR7REFUQUJSSUNLU19IT1NUfSIsIkRBVEFCUklDS1NfVE9LRU4iOiIke0RBVEFCUklDS1NfVE9LRU59IiwiREFUQUJSSUNLU19XQVJFSE9VU0VfSUQiOiIke0RBVEFCUklDS1NfV0FSRUhPVVNFX0lEfSJ9fQ==

Manual Installation

# Clone and enter the repository
git clone https://github.com/markov-kernel/databricks-mcp.git
cd databricks-mcp

# Create an isolated environment (optional but recommended)
uv venv
source .venv/bin/activate  # Linux/Mac
# .\.venv\Scripts\activate  # Windows PowerShell

# Install package and development dependencies
uv pip install -e .
uv pip install -e ".[dev]"

Configuration

Set the following environment variables (or populate .env from .env.example).

export DATABRICKS_HOST="https://your-workspace.databricks.com"
export DATABRICKS_TOKEN="dapiXXXXXXXXXXXXXXXX"
export DATABRICKS_WAREHOUSE_ID="sql_warehouse_12345"  # optional default
export TOOL_TIMEOUT_SECONDS=300
export MAX_CONCURRENT_REQUESTS=8
export HTTP_TIMEOUT_SECONDS=60
export API_MAX_RETRIES=3
export API_RETRY_BACKOFF_SECONDS=0.5

Running the Server

uvx databricks-mcp-server@latest

Tip: append --refresh (e.g., uvx databricks-mcp-server@latest --refresh) to force uv to resolve the latest PyPI release after publishing. Logs are emitted as JSON lines to stderr and persisted to databricks_mcp.log in the working directory.

To adjust logging:

uvx databricks-mcp-server@latest -- --log-level DEBUG

Integrating with MCP Clients

Codex CLI (STDIO)

Register the server and inject credentials via the CLI:

codex mcp add databricks   --env DATABRICKS_HOST="https://your-workspace.databricks.com"   --env DATABRICKS_TOKEN="dapi_XXXXXXXXXXXXXXXX"   --env DATABRICKS_WAREHOUSE_ID="sql_warehouse_12345"   -- uvx databricks-mcp-server@latest
# Add --refresh immediately after a publish to invalidate the uv cache

Or edit ~/.codex/config.toml:

[mcp_servers.databricks]
command = "uvx"
args    = ["databricks-mcp-server@latest"]
env = {
  DATABRICKS_HOST = "https://your-workspace.databricks.com",
  DATABRICKS_TOKEN = "dapi_XXXXXXXXXXXXXXXX",
  DATABRICKS_WAREHOUSE_ID = "sql_warehouse_12345"
}
startup_timeout_sec = 15
tool_timeout_sec    = 300

Planning an HTTP deployment? Codex also supports url = "https://…" plus bearer_token_env_var = "DATABRICKS_TOKEN" or codex mcp login (with experimental_use_rmcp_client = true).

Cursor

{
  "mcpServers": {
    "databricks-mcp-local": {
      "command": "uvx",
      "args": ["databricks-mcp-server@latest"],
      "env": {
        "DATABRICKS_HOST": "https://your-workspace.databricks.com",
        "DATABRICKS_TOKEN": "dapiXXXXXXXXXXXXXXXX",
        "DATABRICKS_WAREHOUSE_ID": "sql_warehouse_12345",
        "RUNNING_VIA_CURSOR_MCP": "true"
      }
    }
  }
}

Restart Cursor after saving and invoke tools as databricks-mcp-local:<tool>.

Claude CLI

claude mcp add databricks-mcp-local   -s user   -e DATABRICKS_HOST="https://your-workspace.databricks.com"   -e DATABRICKS_TOKEN="dapiXXXXXXXXXXXXXXXX"   -e DATABRICKS_WAREHOUSE_ID="sql_warehouse_12345"   -- uvx databricks-mcp-server@latest

Working with Tool Responses

structuredContent carries machine-readable payloads. Large artifacts are returned as resource_link content blocks using URIs like resource://databricks/exports/{id} and can be fetched via the MCP resources API.

result = await session.call_tool("list_clusters", {})
summary = next((block.text for block in result.content if getattr(block, "type", "") == "text"), "")
clusters = (result.structuredContent or {}).get("clusters", [])
resource_links = [block for block in result.content if isinstance(block, dict) and block.get("type") == "resource_link"]

Progress notifications follow MCP’s progress token mechanism; Codex surfaces these messages in the UI while a tool runs.

Example - SQL Query

result = await session.call_tool("execute_sql", {"statement": "SELECT * FROM samples LIMIT 10"})
print(result.content[0].text)
rows = (result.structuredContent or {}).get("result", [])

Example - Workspace File Export

result = await session.call_tool("get_workspace_file_content", {
    "path": "/Users/user@domain.com/report.ipynb",
    "format": "SOURCE"
})
resource_link = next((block for block in result.content if isinstance(block, dict) and block.get("type") == "resource_link"), None)
if resource_link:
    contents = await session.read_resource(resource_link["uri"])

Available Tools

Development Workflow

uv run black databricks_mcp tests
uv run pylint databricks_mcp tests
uv run pytest
uv build
uv publish --token "$PYPI_TOKEN"

Testing

uv run pytest

Pytest suites mock Databricks APIs, providing deterministic structured outputs and transcript tests.

Publishing Builds

Ensure PYPI_TOKEN is available (via .env or environment) before publishing:

uv build
uv publish --token "$PYPI_TOKEN"

Support & Contact

• Maintainer: Olivier Debeuf De Rijcker (olivier@markov.bot)

• Issues: GitHub Issues

• Architecture deep dive: ARCHITECTURE.md

License

Released under the MIT License. See LICENSE.