What is Model Context Protocol (MCP) — and why Salesforce teams should care in 2025

Salesforce development is changing — and fast. With Agentforce 3 ushering in a new generation of AI-powered tooling and smarter DevOps workflows, there’s one term you’re going to hear a lot more about in 2025: MCP. Short for Model Context Protocol, MCP is the behind-the-scenes standard powering many of the new capabilities in Agentforce. In this blog post, guest author Dave Rant (Development Team Lead at Gearset) deep dives into what MCP is — and why it matters for Salesforce professionals.

What is MCP?

Model Context Protocol (MCP) is an open standard, originally developed by the AI safety and research company Anthropic, designed to standardize how AI applications and Large Language Models (LLMs) connect with and utilize external data sources and tools. It’s since seen broad adoption by major players like OpenAI, Google, Microsoft, and Salesforce, positioning it as the de facto standard for agentic interoperability.

The common analogy, “USB-C for AI”, is a useful starting point, conveying the idea of a universal connector that eliminates a tangled mess of proprietary integrations. MCP provides a standardized way to connect AI models to different data sources and tools, overcoming the fragmentation that previously required custom integration code for each tool.

What’s under the hood: MCP architecture

MCP’s architecture follows a classic client-server model, composed of three distinct components that work in concert to connect an LLM to a capability:

• Host: The host is the user-facing application where the AI or LLM is embedded. This could be an AI-powered Integrated Development Environment (IDE) like Cursor, a chat interface like Anthropic’s Claude Desktop, or a large-scale enterprise platform such as Salesforce Agentforce. The host is responsible for managing the user interaction and orchestrating the use of AI.

• Client: The client is a software component that runs within the host. Its sole purpose is to implement the MCP specification and manage the communication with one or more MCP servers. It establishes and maintains dedicated, one-to-one persistent connections to these servers, acting as the protocol-level intermediary between the host’s AI and the external tools.

• Server: An MCP server is a lightweight program that acts as a wrapper or adapter for a specific capability or data source. A server can be built to expose a local file system, provide access to a remote API, wrap a connection to a database, or even represent a complex business application.

The core primitives: tools and resources

An MCP server exposes its capabilities to an AI agent through fundamental primitives. These primitives form the vocabulary that allows an agent to understand what it can know and what it can do.

• Tools: These are the “verbs” of the MCP world. Tools represent functions or actions that the AI agent can execute via the server. Examples include a deploy-metadata tool in a developer context or a create-contact tool for a CRM integration.

• Resources: These are the “nouns”. Resources represent data, documents, or information that the AI agent can access to inform its reasoning and actions. This could be a specific file from a file system, a set of records from a database query, or the response from an external API call.

How MCP powers Agentforce 3

The native MCP client

The most transformative feature of Agentforce 3 is the inclusion of a native MCP client, slated for pilot in the July 2025 release. This component fundamentally alters the nature of the Salesforce platform. It empowers Agentforce agents to connect to any MCP-compliant server, regardless of who built it or where it is hosted, all without writing custom integration code. This move effectively opens the gates of Salesforce’s historically “walled garden” ecosystem, enabling agents to communicate and collaborate across system boundaries to achieve complex goals.

Salesforce MCP toolkit: hosted servers, DX, MuleSoft, and Heroku

Salesforce is actively building a comprehensive, multi-layered toolkit to encourage the creation and management of MCP servers within its ecosystem.

• Salesforce hosted MCP servers (pilot): A no-code/low-code solution that allows administrators to expose specific Salesforce APIs and data objects as fully managed MCP servers, inheriting the platform’s security and trust capabilities.

• Salesforce DX MCP server (developer preview): An open-source, local MCP server designed for pro-code developers. It wraps the Salesforce Developer Experience (DX) command-line tools, exposing capabilities like deploying metadata and running Apex tests as MCP tools.

• MuleSoft Anypoint Platform (general availability): The MuleSoft MCP Connector allows integration teams to convert any existing API or Mule application into a fully governed, agent-ready MCP server.

• Heroku (general availability): For teams that need to build fully custom MCP servers, Heroku provides a managed infrastructure platform with built-in DevOps automation, while Heroku AppLink provides a secure mechanism to connect these servers back into Agentforce.

MCP server registry

Agentforce 3 will introduce an enterprise-grade MCP server registry within the AgentExchange. This centralized registry will serve as the administrative control plane where organizations can discover, approve, and manage all MCP servers that their Agentforce agents are permitted to interact with. Within this registry, administrators can enforce granular security policies, manage identity and authentication, and define rate-limiting and access protocols for each connected server, providing a robust mechanism for governing the use of external tools.

Why it matters for Salesforce DevOps

For Salesforce teams, MCP is set to become the go-to mechanism for integrating third-party systems with AI agents. The strategic integration of MCP into the Salesforce platform dramatically simplifies how agents connect to the wider enterprise technology stack.

The native MCP client in Agentforce 3 provides a “plug-and-play” capability, allowing agents to securely connect to any MCP-compliant third-party system without writing bespoke integration code. This significantly reduces the complexity and time required to extend an agent’s capabilities beyond Salesforce, such as querying a logistics provider’s shipping data in real-time.

For enterprises with existing APIs and legacy systems, the MuleSoft MCP Connector offers a powerful on-ramp to the agentic ecosystem. It allows teams to convert any existing API into a governed, agent-ready MCP server, leveraging current investments while ensuring security and control. For fully custom requirements, Heroku provides a managed platform to build and host MCP servers in any language, with Heroku AppLink ensuring a secure and simple connection back to Agentforce.

This comprehensive toolkit positions MCP as the standard for integrating external systems with Agentforce. By standardizing the connection layer, Salesforce enables teams to build more powerful, cross-system automations faster and more securely than ever before.

Final thoughts

The emergence of Model Context Protocol marks a foundational shift, but it’s only one piece of the puzzle. While MCP standardizes how an individual agent connects to tools and data, the next frontier is enabling agents to collaborate with each other. This is the goal of the Agent-to-Agent (A2A) protocol, an open standard supported by Salesforce, Google, and over 50 other technology partners.

MCP and A2A are complementary, not competing, technologies. Think of it this way: MCP is the protocol that allows a specialized agent to use its tools (the agent-to-tool connection), while A2A is the protocol that allows that specialized agent to talk to other agents (the agent-to-agent connection). With A2A handling how agents communicate and MCP handling how they access the outside world, a cooperative digital workforce can be created.

Salesforce is embracing both open standards to build a truly interoperable ecosystem. Agentforce 3’s native MCP support is the first step, with future support for A2A planned for AgentExchange. The combination of these protocols will allow for complex, multi-system workflows where different agents, each with unique skills, can collaborate to solve business problems autonomously.

Ready to upskill?

If you’re ready to future-proof your Salesforce DevOps knowledge, DevOps Launchpad is here to help. With free courses and certifications designed for Salesforce professionals, you can become an expert in Salesforce DevOps: from automation and deployments to metadata mastery and, yes, even Agentforce fundamentals.