TL;DR Article Summary
Databricks connects all of your data sources across disparate cloud platforms, which already makes it a powerful tool. That power levels up when you employ the Databricks agentic platform and MCP tools. These tools provide natural-language interfaces for your complex data sources. You can build agents to leverage these systems and deploy them on Databricks trivially. Workato’s Enterprise MCP capabilities knock things up another notch. It does so by integrating observability, governance, and compliance tools. Thereby, ensuring that your users are doing their best work in the safest way while working with your MCP servers.
What Is Databricks MCP?
The explosion of LLM systems and agentic workflows has created a problem in many organizations: how should you connect those agents to external data sources? That question has been answered by MCP servers.
MCP stands for Model Context Protocol, and it’s a method of defining ways for LLM agents to interact with a myriad of data sources. MCP.so is a third-party MCP marketplace that defines 22,000+ available MCP integrations.
That enormous list of available integrations raises a new question: how do you manage all of those MCP integrations? That’s where Databricks comes in.
Databricks is not fundamentally an MCP management platform but is rather a platform that’s intended to handle all of your data sources across AWS, GCP, and Azure.
As part of that remit, Databricks provides a control plane that manages MCP servers for your entire enterprise. Databricks handles access control, credential management, and centralized visibility for all of your agentic MCP needs.
In this post, we’ll break down how Databricks works with MCP servers and can integrate into your environment.
How Does Databricks MCP Work?
Databricks defines three different kinds of MCP servers that the platform handles within the broader data management capabilities. Let’s walk through each of them.
Managed MCP Servers
Managed MCP servers work directly out of the box with your existing AWS services. Not coincidentally, this is also the most limited set of MCP servers that you can choose. Databricks provides the following supported managed MCP services:
Genie One
This MCP server allows you to make natural language queries of your Genie Spaces within Databricks. These queries are read-only, meaning you can’t modify data, but Genie One will search and collate data across all of your spaces to answer your query as needed.
Genie Space
Like Genie One, but limited to a single space.
AI Search
This searches across AI vectors to perform searches that will assist your AI agents across many types of files, not just text-based data sources.
Databricks SQL
Run AI-generated SQL to power tools like Codex or Claude Code workflows. Allows both reading and writing data.
Unity Catalog
This catalog stores pre-created SQL statements that AI agents can leverage to perform specified functions.
External MCP Servers
For external servers, things don’t work out of the box. Instead, Databricks serves as a proxy that handles authentication and token management for MCP servers that you’ve installed in your environment.
Authentication and token management are often the two most tedious parts of managing MCP server usage, so Databricks provides a valuable service for enterprises and developers alike.
Custom MCP Servers
Databricks allows you to host custom MCP servers that your business develops. Once you’ve developed and registered a custom MCP server as a Databricks app, you can plug it into Databricks just the same as an existing MCP server, either managed or external.
How Can I Build an AI Agent Using Databricks?
Running AI agents on the Databricks platform enables you to do anything that a normal AI agent can do. Databricks provides different varieties of AI agents that you can create and register:
- Agents that query multiple LLM models side-by-side
- Tool-calling agents
- Retrieval-augmented generation apps
- Multi-agent/multi-step systems
If you’re not code-savvy but you’re still trying to build agents, Databricks has you covered: they have a no-code agent builder which you can try out in sandbox mode.
Manually Building a Databricks Agent
For starters, if you’re building an agent that’s simply coordinating between tool calls and doesn’t expose any custom logic between those calls, Databricks provides a Supervisor API that enables building agents for those workflows without requiring you to write custom code.
If you’re wrapping custom code around tool calls, then Databricks supports that, too.
Architecture showing an AI agent using MLflow, an LLM, and MCP servers to access external resources.
The Databricks platform allows you to develop and evaluate agents using Python, and it supports agents written using any authoring library.
Practical Recommendations for Your Agents
Databricks provides a very wide base for building and deploying agents. As you build them, here are some practical recommendations for making sure that you’re building the best agents for your organization.
Use All of the Tools
Many teams will decide that they need to jump immediately to developing custom, complicated code for their agents. In reality, no-code or low-code options on Databricks will often get the job done with much less fuss.
Keep it Simple, and Limit the Scope
The common mental image of agents is that they’re large, all-purpose tools. But in reality, the most effective agents are tightly focused tools that do one thing and do it well.
Databricks provides you with many options for proliferating agents, meaning that there’s minimal cost for building many small, focused agents.
Iterate Often
Shipping new versions of your Databricks agents is simple and straightforward, regardless of whether you’re writing custom code or using a no-code option.
So, ship your small, focused agents and then tweak them as you evaluate their results, so that you’re constantly providing your users with the best agentic options at all times.
How Does Workato Extend Databricks MCP Enterprise Workflows?
Workato offers Enterprise MCP tools, which present another level for your MCP server integrations. And of course, those tools integrate directly with Databricks, meaning that you can feel confident in the work that your users are doing.
That work will be safe and secure due to Workato’s tools. Let’s look at how Workato extends your MCP integration functionality.
Observability
Workato offers full observability into every agentic interaction for your users. Whether you need that observability for accountability or debugging less-than-ideal outcomes, the data about your agentic interactions is at your fingertips.
Governance
One of the most common security concerns about the proliferation of AI agents is that users will start adopting “Shadow AI,” creating their own integrations that put their data at risk of deletion or leakage.
Workato’s governance features provide you with the security that the actions taken by your users are only those they’ve been approved to take.
Security
Enterprise-grade authentication and authorization techniques ensure that every agent runs securely by default. Your organization can rest easy understanding that your agents won’t do anything that they shouldn’t.
Compliance
All of these features add up to world-class compliance capabilities that mean that your agentic workflows comfortably comply with frameworks like SOC II, PCI, and GDPR.
Workato and Databricks Work Hand-in-Hand to Make Your MCP Agents Shine
Databricks provides unparalleled flexibility for connecting agents to your data sources, no matter which cloud they live in.
Workato makes those agents safe, secure, and more effective through their governance tools, and they offer you improved troubleshooting from an outstanding suite of observability tools.
When those forces combine, the result is agent-enabled workflows that make work easier, faster, and safer for all your employees. If you’re intrigued by how these tools can work together for you, Workato would love to provide a demo of what we can do for you.
FAQ
What Are LLM Agents
LLM agents are code-based harnesses for AI models that provide those models with the capability to interface with software-based tools in their environment.
The concept of an “agent” is somewhat overloaded at the moment, where an “agent” both means a simple tool-using LLM harness and something gargantuan and all-purpose like OpenAI Workspace agents or Claude Code.
The thing that connects all these different concepts of software is their ability to use tools to respond to queries from the user.
Should My Company Build its Own MCP Servers?
The decision whether or not to build MCP servers for internal systems for your company is a complicated one, with many factors to consider. MCP servers provide network-based access to software tools on your internal systems, but they aren’t required for agent-based tool access.
LLM agents are fully capable of interacting with traditional HTTP-based APIs, so you may not need a full MCP server. However, whichever approach you choose, you’ll still need to answer questions about things like authentication and authorization, governance, and observability.
What Is the Best MCP Server?
In reality, there’s no such thing as a “best MCP server.” Adding a Jira MCP server to your environment is great…if you use Jira. If you don’t, it’s useless! Asking for the best MCP server is a bit like asking for the best programming language or the best screwdriver.
In reality, they’re specialized tools, and identifying the best one is about understanding the problem you’re trying to solve and how each tool goes about doing that.
This post was written by Eric Boersma. Eric is a software developer and development manager who’s done everything from IT security in pharmaceuticals to writing intelligence software for the US government to building international development teams for non-profits.