In a world where AI agents are becoming the frontline interface for customers, sales reps, and service teams, there’s one technical challenge that keeps surfacing: context management. How can an AI agent consistently retrieve the right data, perform the correct actions, and stay in sync across tools and systems — all in real time?
Enter the Model Context Protocol (MCP) — a new open standard that Salesforce has now embedded into Agentforce 3, allowing AI agents to interact with external systems in a secure, structured, and real-time way.
What Is Model Context Protocol (MCP)?
MCP is an open protocol originally developed by Anthropic that defines how AI agents can access tools, fetch external data, and reuse prompts in a standardized format. Think of it like USB-C for AI — once your data or function is MCP-compliant, it can be plugged into any supported AI system.
As of mid-2025, Salesforce has officially adopted MCP inside Agentforce 3, turning it into a native client. That means your Salesforce AI agents — powered by Einstein or custom LLMs — can now seamlessly query tools, retrieve live data, and execute actions from any MCP server, without custom wiring.
Why MCP Matters in Agentforce
Without a structured protocol like MCP, AI agents often rely on:
- Custom prompt injections
- Hardcoded tool access logic
- Repetitive context packing per use case
This leads to brittle, non-reusable agent logic — and worse, context fragmentation across teams or workflows.
By contrast, MCP enables:
- Reusable tools and APIs: Register once, use anywhere in Agentforce
- Standard context schemas: No guesswork for agents interpreting external data
- Governed access: All MCP tools and resources go through Salesforce’s security model
This makes MCP ideal for real-time applications, especially when agents need to reason, act, and respond within a single session.
How MCP Works in Agentforce
Core Components:
Host – UI where conversation happens (Chat, Slack, Console, etc.) in my case Github Copilot or Agentforce 4 Developers
MCP Client – Agentforce’s internal handler for all MCP calls
MCP Server – External service exposing resources, tools, or prompts
MCP Server Types:
- Resources – Read-only data (e.g., shipping status, customer history)
- Tools – Callable actions (e.g., create case, send invoice)
- Prompts/Templates – Reusable prompt blocks for consistency (e.g., refund policy wording)
These are registered through the Agent Gateway and managed via admin-approved scopes and trust policies.
Real-Time Use Case Example
Let’s walk through a customer support scenario:
A customer asks in chat: “Can I get an update on my return?”
- Agentforce passes the customer message and metadata to the LLM.
- The agent makes an MCP resource call to a logistics server, fetching the return status.
- If needed, the agent invokes an MCP tool to email an update or trigger a refund process.
- The final response is crafted using a prompt template that aligns with your brand tone.
- All of this happens without a page reload, API bounce, or manual agent involvement.
That’s real-time, multi-system reasoning — powered by MCP.
Setting Up MCP in Agentforce
Check my How to Get Started Using the Latest Salesforce DX MCP Server youtube video and my post
Best Practices
- Start with Resources, then Tools: Let agents observe data before acting.
- Centralize Prompts: Keep tone and language consistent across departments.
- Rate Limit Everything: Prevent infinite tool loops or agent misuse.
- Use Real-Time Logging: Validate and debug behavior in sandbox before going live.
Final Thoughts
Model Context Protocol is more than just a buzzword — it’s the backbone of real-time, secure, and scalable AI agent systems. In Salesforce Agentforce, MCP transforms your org’s tools, data, and prompts into first-class capabilities your AI agents can reason with, act upon, and respond through.
If you’ve already embraced Agentforce, MCP is the natural next step. And if you’re just getting started, begin small — one tool, one prompt, one use case — and build from there.