What is MCP and how do startups use it to connect agents to tools?

Summary

• The Model Context Protocol (MCP) is an open-source standard that lets AI agents dynamically discover and invoke external tools at runtime without hardcoding integrations.
• Startups use MCP to connect agents to enterprise systems, databases, SaaS applications, and internal services through a single, standardized protocol.
• Databricks provides three server types for MCP: Managed servers for Databricks data sources (Unity Catalog Functions, Genie Agents, AI Search), External servers for third-party services, and Custom servers hosted as Databricks Apps.
• MCP integrates with governance layers (Unity Catalog, Unity Gateway) so tool access respects security policies and user permissions automatically.
• Supported agent frameworks include LangGraph, CrewAI, and the Claude Code SDK, letting teams build with tools they already know.

What is MCP and how do startups use it to connect agents to tools?

The Model Context Protocol (MCP) is an open-source standard that enables AI agents to discover and interact with external tools and data sources at runtime. Instead of hardcoding connections to each service, MCP allows AI agents to dynamically discover available tools, understand their capabilities through structured calls, and invoke them with proper governance and permissions. This standardized approach reduces boilerplate integration code and lets startups build agents that work with diverse enterprise systems without writing custom connectors for each one.

Why MCP matters for agent builders

Building AI agents typically requires integrating with many external systems: databases, APIs, SaaS platforms, internal tools. Without a standard protocol, developers either hardcode integrations (inflexible and brittle) or build custom adapters (expensive and repetitive). MCP solves this by providing a common language for agents and tools to communicate, enabling developers to focus on agent logic rather than plumbing.

Three server types for connecting agents to tools

Databricks provides multiple ways to connect agents to tools using MCP:

Server Type	Best For	Governance	Example
Managed MCP servers	Databricks data sources (Unity Catalog Functions, Genie Agents, AI Search Indexes)	Built-in; respects UC permissions and user roles	Querying AI Search Indexes with user-scoped results
External MCP servers	Third-party services and SaaS platforms	Registered as governed MCP Services in Unity Catalog	Connecting to SaaS CRM, ERP, or analytics tools
Custom MCP servers	Internal services and proprietary tools	Hosted as Databricks Apps; OAuth support	Custom business logic, internal microservices

How startups build agents with MCP

No-code prototyping. Use the AI Playground to prototype agents with MCP tools without writing code. Select any tools-enabled LLM and use the "Tools" dropdown to connect managed or custom servers.
Agent Bricks integration. The Supervisor Agent in Agent Bricks supports MCP servers built as Databricks Apps, letting you add internal services with a single click.
Framework flexibility. Agent Bricks supports popular open-source agent harnesses including LangGraph, CrewAI, Agno, and the Claude Code SDK, so teams can build with frameworks they already know.
Connection methods. Developers can connect agents to tools through:

• Managed MCP servers for Databricks data sources
• External MCP servers registered as MCP Services
• Unity Catalog functions for custom logic
• Custom tools built with the Agent SDK

How MCP works with LLMs

MCP enables LLMs to take actions more naturally by giving them the context they need. The protocol works in three steps:

1. Tool Discovery: LLMs dynamically discover available tools at runtime rather than relying on hardcoded integrations. This allows agents to adapt as new tools are added.
2. Structured Calls: Tools are understood through structured schemas that define their capabilities, parameters, and return types.
3. Execution with Permissions: Tool invocation respects governance and security controls, ensuring users can only access tools and data they are authorized to use.

Governance integration. Databricks combines MCP with Unity Catalog and Unity Gateway to provide governance, security, and discoverability. Managed servers support on-behalf-of-user authentication out of the box, respecting permissions established in Unity Catalog. OAuth is supported as standard in Databricks Apps for easy hosting.
Observability. When deploying agents on Databricks, you can see which tools are being used and how through MLflow and Agent Evaluation, helping you optimize agent behavior and tool selection.

Common startup use cases

Customer support automation. Startups connect agents to CRM and knowledge systems to provide instant, personalized support without human intervention.
Internal operations. Connect agents to payroll, expense, and project management systems to automate routine tasks and free up operational teams.
Data analysis and reporting. Agents query data warehouses and analytics systems to generate insights and answer business questions on demand.
Multi-tool workflows. Agents coordinate across multiple systems (booking, inventory, finance) to orchestrate complex business processes.

Getting started with MCP on Databricks

For technical setup and code examples, see the Databricks agents documentation and the MCP overview blog post.

FAQs

What's the difference between MCP servers and Unity Catalog functions?

Both can be used as agent tools. MCP servers are stateful services that can maintain connections or context; Unity Catalog functions are simpler, stateless operations. Use functions for straightforward transformations; use MCP servers for services that need to manage state or complex authentication.

Can I use MCP with open-source LLMs?

Yes. MCP is a protocol-level abstraction, so it works with any LLM that supports tool calling, including open-source models. Databricks supports LLMs from Anthropic, OpenAI, Meta, and others through the Model Serving platform.

How secure is MCP?

MCP itself is just a protocol. Security depends on implementation: managed servers inherit Databricks' governance (Unity Catalog, Unity Gateway); external servers should use OAuth or other standard auth mechanisms; custom servers should follow security best practices (authentication, encryption, least privilege).

Do I need to write code to use MCP?

No. You can prototype with MCP in the AI Playground or Agent Bricks without code. For production deployments, developers typically write code to configure servers and define tool logic, but Databricks provides SDKs and templates to simplify this.

Can agents use multiple MCP servers?

Yes. Agents can discover and use tools from multiple MCP servers in a single conversation, enabling complex workflows that touch many systems.