More and more teams are connecting AI tools to Salesforce. That means more integrations to build. More to secure. More to maintain.
And every time a new AI tool shows up, someone on the team has to figure out how to make it talk to the CRM without breaking what's already there. It gets messy, fast.
Salesforce saw this coming and started rolling out support for something called the Model Context Protocol across its platform. You might have seen it called MCP.
If you're wondering what it is and whether your org needs it, this is a good place to start.
MCP stands for Model Context Protocol. It's an open-source standard created by Anthropic and now supported by Google, Microsoft, AWS, and Salesforce under the Linux Foundation.
In simple terms, it gives AI agents one shared way to connect with your Salesforce data.
Here's why that matters.
Say your team uses Claude for account research, Cursor for development, and Agentforce for customer service. Without MCP for Salesforce, each of those tools needs its own connector to talk to your CRM. That's three integrations to build. Three to secure. Three to fix when something breaks. Add another tool next quarter and you're starting from scratch again.
MCP changes that. Your org exposes its data through one MCP server. Any AI tool that supports the protocol can plug into it. One server, many tools. No more rebuilding the same connection over and over.
This is part of a bigger shift Salesforce calls Headless 360. The idea is simple. Every workflow, every record, every process in your org becomes something an AI agent can access directly. No screen, no clicks, just a clean connection through MCP.
A quick look at the moving parts behind Salesforce MCP and how they fit together.
MCP has three parts:
These three parts work together every time you ask a question or trigger an action through AI.
Say you type: "Which accounts have a renewal coming up but no activity this month?"
Here's what happens behind the scenes:
One thing to remember here: the AI never touches your Salesforce data directly. It only sees what the MCP server allows.
Now that you know how the request moves, it helps to understand what exactly the server is exposing to the AI in the first place.
MCP servers share three types of capabilities with AI agents:
Salesforce doesn't offer just one MCP server. There are several, each built for a different use case. Here's what's available today and who each one is for.
These are cloud-managed servers that Salesforce runs for you. No local setup, no infrastructure to maintain. You authenticate using OAuth 2.0 and your AI tool connects directly to Salesforce's endpoint.
What makes Salesforce hosted MCP servers stand out is how they're split by risk level:
On top of these, there are servers for specific Salesforce products:
This tiered approach lets admins control exactly how much access AI agents get, without building custom permission logic.
This one is built for developers. It runs locally through npx @salesforce/mcp and connects using your existing Salesforce CLI credentials.
It comes with over 60 tools spread across 15 toolsets, including:
You pick which toolsets to load using the --toolsets flag. This matters because there's a practical limit of around 20 tools running at the same time before things start slowing down.
It's open source under Apache 2.0 and already has over 121,000 monthly npm downloads.
Data 360 has close to 200 API operations. Exposing all of them as individual tools would overwhelm any AI model's context window. So Salesforce built something smarter.
This server uses a facade architecture with just three tools:
Three tools on the surface, 190 operations underneath. It keeps things clean for the AI while still giving you access to ingestion, identity resolution, segmentation, transforms, and more.
Currently requires Java 17+ and Maven 3.9+. The GA version will be a fully hosted server with no local setup needed.
Already have REST or SOAP APIs running in MuleSoft? This connector turns them into MCP servers using configuration alone. No new code needed.
It automatically adds security policies, activity tracing, and rate limiting. The AI Agent Gateway on top acts as a centralized registry where admins can manage which MCP servers are allowed, who can access them, and what limits apply.
This is the fastest path for enterprises that want to make their existing integrations AI-ready.
For teams that need something fully custom. You write your own MCP server in any language, deploy it on Heroku with a standard Git push, and connect it to Agentforce through Heroku AppLink.
Heroku handles the infrastructure, scaling, and DevOps. You focus on the logic.
This is the other side of the equation. Instead of serving data, Agentforce consumes it. It has a native MCP client that can connect to any third-party MCP server without custom code.
MCP actions show up like regular agent actions inside Agent Builder. Admins govern which external servers are allowed through the MCP Server Registry in AgentExchange.
Security follows the same rules as the rest of Salesforce. Every MCP tool call runs with the same permissions as the authenticated user. Field-level security, object access, and sharing rules all apply automatically. If a user can't see a field in Lightning, the MCP server can't return it either.
Getting started with Salesforce MCP depends on which server type you're planning to use. Here's what you need to know.
Salesforce MCP is available on:
If you're on Salesforce Foundations, you also get 200,000 Flex Credits included. That's enough to pilot and test before going org-wide.
Not every AI tool supports MCP yet. Here's what works with Salesforce MCP today:
Any client that supports OAuth 2.0 Authorization Code with PKCE will work. For clients that only support stdio transport, use the mcp-remote package to bridge the connection.
This is the quickest way to get hands-on. You'll need:
The Salesforce MCP setup takes just a few minutes:
No OAuth needed. It runs on your existing CLI credentials.
Hosted servers need a bit more configuration, but they're built for org-wide use.
AgentExchange is the place to start. It now lists over 1,000 agents, tools, and MCP servers from Salesforce and its partners. You can browse, connect, and deploy them through Agent Builder with no code required.
There's no separate license for MCP. You pay based on usage through Flex Credits.
Quick example: 20 reps running 10 queries a day, 3 actions per query, over 20 working days comes to about 12,000 actions a month. That's roughly $1,200.
Three things trip people up during setup more than anything else:
Every platform shift comes with trade-offs. Here's an honest look at what Salesforce MCP gets right and where it still falls short.
Before MCP, every AI tool that needed Salesforce data required its own custom connector. That meant separate builds, separate maintenance, and separate troubleshooting for each one. With MCP, you build one server and any AI client that supports the protocol can connect to it. Add a new tool next month and it works out of the box.
Every MCP tool call runs with the same permissions as the authenticated user. Field-level security, object access, sharing rule, all enforced automatically. If a user can't see a field in Lightning, the MCP server can't return it either. You don't need to build a separate permission layer for AI access.
This is the one most teams overlook. Every Flow, every Apex class, every MuleSoft integration you've already built can now be exposed to AI agents through MCP. You're not starting from scratch. Years of customization become usable in a completely new way without rewriting anything.
A sales manager can ask "Which renewals are at risk this quarter?" and get an answer in plain language. No SOQL query. No report builder. No waiting for someone else to pull the data. The AI handles the query and returns what they need, right where they're already working.
MCP is an open standard under the Linux Foundation. Your servers work with Claude, ChatGPT, Cursor, and any other compliant client. If you switch AI tools tomorrow, your MCP servers still work. There's no vendor lock-in at the protocol level.
That said, there are real gaps worth knowing about before you go all in.
There's a practical limit of around 20 tools running at the same time. Go beyond that and the AI model's context window gets overloaded. It starts picking the wrong tool, making up arguments, or losing track of what you originally asked. Choosing the right set of tools for each use case is an operational decision, not a set-it-and-forget-it task.
MCP returns raw field names and values. It doesn't return the context behind them. It won't know why that Opportunity was marked Closed Lost, or that your team stopped using a particular field two years ago. This gap between data and meaning is where most "confidently wrong" answers come from.
Every MCP tool call translates into one or more Salesforce API calls under the hood. Enterprise Edition starts at 100,000 requests per 24 hours. That sounds like a lot until you have multiple agents running across multiple teams, each making several calls per query. Plan your API budget before scaling.
The native MCP client in Agentforce, the Data 360 MCP Server, and several DX toolsets are not yet generally available. Beta means no SLA, possible breaking changes, and no guaranteed GA timeline. That's fine for testing and pilots. It's a risk for production workflows.
Make your tools too specific and the AI can't use them on their own. Each tool returns a partial result that only makes sense when called in a certain order, and the AI doesn't know that order. Bundle too much into one tool and the AI can't reason about what's happening between steps.
MCP is still early. But the direction is clear. Salesforce is making every part of its platform accessible to AI agents through a single protocol, and the ecosystem around it is growing fast.
The teams that get ahead with MCP won't be the ones who enabled it first. They'll be the ones who took time to set it up right. Choose the right server types. Structured their tools properly. Planned for API limits before they hit them.
That's the kind of work we do at MIDCAI. As a Salesforce consulting partner, we help teams go from "we've heard about MCP" to "it's running in production and our agents actually return answers we trust." If that's where you're headed, we'd love to help.
5+ years of experience across content, social media, and growth marketing. At MIDCAI, I focus on turning complex ideas around AI, data, and technology into clear insights and campaigns that help businesses attract and engage the right audience.
Get in touch with us for any enquiries and questions
Define your goals and identify areas where technology can add value to your business
We are looking for passionate people to join us on our mission.
where your skills fuel innovation and your growth powers ours
