The Protocol That Every AI Company Is Quietly Building On | What is MCP? Ft. Marc Benioff

I recently sat down with Marc Benioff and at the time they had just announced 30 new capabilities for Slackbot all in a single keynote and he told me something that I haven't really been able to stop thinking about.

The user interface to the whole AI ecosystem. I mean that's a massive architectural claim and Benioff isn't the only one making it. I mean Microsoft is building the same argument for say Teams with Copilot. Google is doing it with Gemini inside every workspace. Basically every major platform company is racing to become the single service that every model, every agent runs through. But here's the thing. The reason Slack might actually be able to pull this off has nothing to do with Salesforce's sales team. It has to do with a protocol that most people outside of engineering have never heard of. And that protocol is quietly becoming the most important piece of infrastructure in the entire AI stack. Okay. Back in November of 2024, Anthropic open-sourced something called the Model Context Protocol, MCP. The simplest way to understand it is this. Before MCP, every time you wanted an AI model to talk to an external tool, you had to write a custom integration. For example, if you want Claude to query your Postgres database, custom code. You want it to also pull from Slack, different custom code. You want it to work with say Jira, another integration. Every combination of model and tool required its own bespoke wiring. MCP standardizes all of that. One protocol, any AI model that speaks MCP can connect to any tool that has an MCP server. The analogy people keep using is USB-C. Before USB-C, every device had its own charger, its own cable, its own connector. MCP is doing the same thing for AI agents. One standardized interface between the model and the world. By early this year, MCP had over 97 million monthly SDK downloads. 97 million. More than 500 public MCP servers. Official SDKs in TypeScript, Python, C, Java, and OpenAI adopted it across ChatGPT. Google supports it. Microsoft supports it. Anthropic donated it to Linux Foundation. In about 18 months, it went from an open-source side project to the connective tissue of the entire agentic AI stack. Now, when I was talking with Marc, he kept coming back to that one idea that the pattern he sees across every major tech wave, whether it be cloud, mobile, AI, is the same.

That is exactly what MCP enables and it explains why Salesforce is betting everything on Slack as a surface layer. Okay. In February 2026, Slack launched its own MCP server. That meant any external AI client, Claude, Perplexity, OpenAI, could pull data from Slack and take actions inside of it. Then in March, Slack became an MCP client too. Meaning Slackbot itself can now reach outward and call into any external tool or service that runs an MCP server. And that's bidirectional at its finest. Agents can talk to Slack. Slack can talk to agents. And because MCP is an open standard, Salesforce didn't have to build custom integrations with each AI company individually. They basically built one MCP implementation and every model that speaks the protocol can plug into it. Now Marc told me he does basically all of his work in Slack.

And I experienced the same thing before the keynote. I was going through the Slackbot demos and I realized something. I'm a big user of Perplexity's computer and then I'm also a user of Claude Code. And when I was going through these examples today before the keynote, I thought I no longer have to go to any of these other interfaces. I can do it all right in Slackbot. That experience I described using Perplexity and Claude and other tools without leaving Slack, that only works because of MCP. So here's what's actually happening under the hood. Slack sits at the top of the stack as an interface layer. Below that is Agentforce, Salesforce's agent orchestration layer. Below that are the applications like Sales Cloud and Service Cloud. Below that is what Salesforce calls Data 360. That's a data layer that pulls from your CRM, your email, your files, your information pipelines, and it harmonizes it so agents can actually understand your business context. Now Marc described this stack during our conversation. So take a listen.

That data layer is the part that most people skip over and it's the part that actually matters for engineers. MCP gives agents a way to connect to tools, but the quality of what those agents can do depends entirely on the data underneath it. When Marc says federated but harmonized, he's basically describing a genuinely hard engineering problem. Enterprise data lives in dozens of different systems with completely different schemas. Your CRM stores customer records one way. Your support tickets are structured differently. Then you have things like your email threads that have no schema at all. For an agent to do useful work across all of those different systems, something has to reconcile those schemas in real time. And that's what Salesforce is building with Data 360. It's basically a translation layer that lets agents reason across data sources that were never designed to talk to each other. And MCP is the protocol that lets the agent actually reach those sources. I mean, Anthropic is using this full stack right now. And at the keynote, Salesforce announced the results. So, Anthropic accelerated their deals. It was announced today at the keynote by 60% using Slackbot.

It's incredible. 60%. And I had to look at the slide twice. That 60% number is what makes this interesting beyond just Salesforce because every platform company is making the same architectural bet right now. Microsoft building Copilot Studio as their agent orchestration layer with Teams as the surface. Going back to Google, Google has Gemini wired into Workspace with their own agent framework. Even OpenAI launched Frontier with the vision where they own the orchestration layer and SaaS companies become data providers underneath. Each one really has a different answer to the same question. Who owns the interface between humans and AI agents? Now, Slack's argument is that a communication tool people already live in with MCP as the protocol layer beats a document editor or a standalone AI chat window. Custom AI agents inside Slack have grown 300% since January. And starting this summer, every new Salesforce customer automatically gets Slack bundled in. To me, and hopefully to you, that really tells you where Salesforce thinks the future of enterprise software lives. Marc told me he thinks the next wave goes beyond language models entirely.

Multisensory models processing data from multiple sources all at once. That means MCP servers won't just be connecting to databases and messaging apps. They'll need to handle sensor data, video feeds, audio streams. I mean, even things like location data. The protocol was designed to be extensible and that extensibility is about to really get tested. Now for engineers, this is where the opportunity is right now. MCP servers once again are open source. The spec is public. Anyone can build one and the companies that build the best MCP integrations for their particular domain are going to have a real advantage as every AI platform competes really to be the one interface. I'm very passionate about this. So Marc told me about a conversation he had at MIT with a computer science student who was thinking about changing her major and shifting gears a little bit, but it's very fascinating. So I really wanted to include this part in the video.

He's right that engineers will be more productive, but I would even go further. Engineers who understand how protocols like MCP actually work, who can build the connective layer between the AI models and the real systems. Those are the engineers who are going to be the most valuable in the next 5 years. The race to become the one interface for AI isn't going to be won by whoever has the best chatbot. It's going to be won by whoever has the best protocol layer underneath it. And right now, MCP is that protocol. Marc said it was about the stack, not just the Slack.

He's more right about that than I think anyone really realizes. All right. And I hope you enjoyed this video. It was so much fun sitting down and having a really candid conversation with Marc and especially looking ahead what the future holds for AI and what it means to have that foundational layer with MCP. I think it's something that not enough people are having conversations around. Leave in the comments who else would you like me to bring on the channel for conversations around where tech is headed.