Connect your AI Agents to Databricks in minutes

What is a Databricks MCP server?

It’s an MCP server that exposes data and functionality from a Databricks instance via tools. Your agents can invoke these tools to create new clusters, list job runs in a given workspace, delete a job, and more.

Databricks offers a suite of “official” MCP servers:

• Vector Search retrieves relevant documents from vector indexes 
• Genie queries structured data via natural language 
• Databricks SQL executes AI-generated SQL for data operations 
• Unity Catalog Functions run predefined SQL functions as tools

You can also use a hosted MCP platform that offers dozens of tools out of the box, such as Merge Agent Handler.

How can I use the Databricks MCP server?

Here are some popular use cases:

• Natural-language analytics over Lakehouse data: Ask questions in plain English, translate into safe SQL/Spark, run in Databricks, and return results with pointers to the underlying tables/views
  

• Data quality monitoring and incident automation: Run scheduled freshness/volume/schema checks, detect anomalies, and automatically open an incident or ticket with the failing queries and impact summary
  

• Pipeline debugging and root-cause analysis: When jobs fail or metrics drift, inspect recent runs/logs and propose likely causes and next-step fixes
  

• Cost and performance optimization: Identify expensive queries and inefficient cluster usage, then recommend prioritized changes (e.g., query rewrites, partitioning strategy, cluster sizing) with measurable savings targets
  

• Governed self-serve data discovery and access: Help teams find the right datasets, explain definitions, and streamline access requests/approvals while keeping an audit trail

What makes Merge Agent Handler’s Databricks MCP server better than alternative Databricks MCP servers?

Here are just a few reasons:

• Enterprise-grade connector infrastructure: Merge Agent Handler provides fully managed connectors that handle authentication, rate limiting, error handling, and lifecycle management so developers can focus on building agent logic instead of maintaining integrations

• Built-in security gateway with DLP controls: Agent Handler includes a Security Gateway that scans tool inputs and outputs for sensitive data and enforces security policies. Teams can configure default or custom rules to block, redact, or mask data before it reaches an agent or external system.
  

• Flexible authentication and credential management: Authentication is handled directly by the platform through registered users and connector authentication flows. And if a user hasn’t authenticated a connector yet, Agent Handler signals the application to prompt authentication via Merge Link

• Real-time observability and searchable audit logs: Agent Handler logs every tool call and underlying API request so teams can monitor agent activity, debug issues, and track system behavior. These logs are fully searchable and form a complete audit trail for compliance, debugging, and operational monitoring

• Customizable connectors and Tool Packs: With Connector Studio, teams can combine Databricks with other enterprise connectors, modify tool definitions, or import additional MCP servers. This enables building Tool Packs tailored to specific agent workflows rather than relying on a single-purpose Databricks MCP implementation

Can I set custom security rules for Databricks tool calls in Merge Agent Handler?

Yes. Merge’s platform materials say Agent Handler includes default security rules and custom security rules, and the Security Gateway scans tool inputs and responses for sensitive data before taking action based on platform configuration. 

Here are a few examples:

• Block PII patterns (inputs + outputs): Block any Databricks tool call if payloads contain SSNs or credit card numbers (regex-based)
  

• Block destructive SQL / admin operations: Block queries containing <code class="blog_inline-code">DROP</code>, <code class="blog_inline-code">TRUNCATE</code>, <code class="blog_inline-code">GRANT</code>, <code class="blog_inline-code">ALTER</code>, or <code class="blog_inline-code">DELETE</code> without a <code class="blog_inline-code">WHERE</code>
  

• Prod boundary rule: Block calls that target prod workspaces/catalogs/schemas unless the Tool Pack (or Registered User group) is explicitly allowed
  

• Sensitive table/column allowlist: Block queries that touch specified tables or fields (e.g., salary) unless running under a privileged Tool Pack
  

• Redact secrets: Allow execution, but redact API keys, bearer tokens, and connection strings from requests/logs to prevent credential leakage

How can I start using Merge Agent Handler’s Databricks MCP server?

You can follow these steps:

1. Create an Agent Handler account and grab your API key from the Agent Handler dashboard. This API key is what your MCP client uses to authenticate to the Agent Handler MCP endpoint.

2. Create a Tool Pack for the Databricks workflow (e.g., “Databricks Analytics”). Tool Packs define which connectors and tools your agent is allowed to call.

3. Create (or select) a Registered User that your agent will act “as” when making tool calls. Registered Users are the identity containers that hold credentials (user-level or shared auth).

4. Authenticate Databricks for that Registered User (or configure shared auth if you want one org-level Databricks credential reused across users). In practice, this is typically done via Agent Handler Link in your product so end users can connect without you handling credentials directly.

5. Connect your MCP client to the Agent Handler MCP endpoint for that Tool Pack + Registered User. Then your client can <code class="blog_inline-code">list_tools</code> and <code class="blog_inline-code">call_tool</code> for the Databricks tools you enabled in the Tool Pack.

6. Run a few real prompts and monitor the results using Agent Handler’s logs (and add rules/alerts as needed) to verify the tool calls and outputs look right before pushing the connector to production.