Agents, MCPs and The Next Billion Dollar Idea

I'm about to share an idea that I believe will be worth billions.

If I wasn't already knee-deep in building agent.ai, I'd probably try to build this myself. It's that compelling.

The opportunity centers around MCP (Model Context Protocol), which is quietly transforming AI and agent capabilities while most people aren't looking.

So, in today's newsletter, I'm breaking down:

• What MCP is in simple terms

• The big problem with MCP today

• My billion-dollar idea: the infrastructure for AI connectivity

—Dharmesh

What's MCP and Why Should You Care?

MCP (Model Context Protocol) is an open standard that's currently on fire (🔥) in terms of adoption among AI early adopters.

In simple terms, it's a protocol that lets AI applications (called “MCP Clients“) connect to “MCP Servers”. You can think of these servers as providing a set of "tools" that the AI application has access to.

By making a standard protocol, any MCP Client can make use of the hundreds and thousands of MCP servers that will be out there. This adds an immense amount of power to tools like Claude, ChatGPT -- and of course agent.ai.

This kind of open standard being widely adopted is what allows really fast innovation and brilliant breakthroughs.

Of course, I’ve been testing MCP extensively. Right now, I have Claude Desktop configured to interact with several MCP Servers from different companies. This gives the LLM hundreds of tools that it can decide to use based on what I enter for a prompt:

• Use agents on Agent.ai

• Access CRM data in HubSpot

• Read/write to a specified directory in my local file system

• Read/write messages to Slack

• Access my Google Calendar and Gmail

The beauty of MCP is that Clients don't need to be custom-coded for specific APIs, and Servers don't need to account for different types of Clients. Everything communicates through a standardized protocol.

Here's an example prompt I recently used: "Lookup OpenAI in the HubSpot CRM and Slack the details to @dharmesh including how long ago I had the last interaction."

I could have done something much more complicated and hit a dozen different systems, but you get the idea.

The Big Problem with MCP Today

Despite all this potential, there's one big problem standing in the way of mainstream adoption:

Right now, finding the right MCP Servers and plugging them into something a chat app like Claude Desktop is messy and scary. It's the wild, wild-west out there. Most of the servers are shared as a GitHub repo and you'd have to self-host them to use them. Ick!

I love the idea of MCP, but this implementation challenge breaks down into several key issues:

• Authentication: How will you determine who has access to what capabilities?

• Trust: How do you determine which MCP servers are trustworthy enough to use?

• Provisioning: MCP servers are currently often offered as a GitHub repo, but you're responsible for self-hosting the code.

• Security: Because of the way LLMs interact with tools made available by an MCP server, there are some new risks. That's why it's important to use known clients with trusted servers.

We (as an industry) are making progress along all of these fronts, but it's important to know that today, MCP is for geeky, early adopters -- or specific use cases that involve specific tools/resources.

The promise of MCP is a large pool of arbitrary clients being able to connect to a large pool of arbitrary servers, creating an exponential explosion of what can be done with AI (but we're quite not there yet).

So what would it take to solve these challenges and bring MCP to the mainstream? This is where I see the billion-dollar opportunity...

The Billion Dollar Idea: MCP.net

So, here's the idea: Build a centralized network of MCP Servers that makes it frictionless to get going and provide fast time to joy. I'd call it MCP.net.

But the network would be more than just a directory of MCP servers. It would allow:

• Anyone to submit an MCP Server and have it hosted on the network

• Ratings/reviews of MCP servers

• Semantic search to find the right tools (embedded within the server)

• A way for users to request access to specific functionality that is not yet on the network

• Creating server "remixes" that combine tools from different servers

• and much, much more...

Think of it as the Hugging Face of MCP. A way to discover and connect to MCP Servers.

I know there are already a few emerging directories out there, and Anthropic is working on a "registry", but I'm thinking about something well beyond that.

Now you're wondering why I don't just go do this. Couldn't I just position it as a "Professional Network for MCP Servers" (playing off the positioning of Agent.ai)? I could probably Vibe Code the first 50%. 😀

But focus matters. I'm already building agent.ai, and dividing my attention would be a mistake. I'm hoping someone with street cred is already doing this. I hope so. That way, I can focus on the thing I should be doing.

I actually posted about this idea on LinkedIn last week and ended up investing in a founder who reached out through the comments! That's how much I believe in the potential of this space.

I don't usually share business ideas I'm not pursuing, but this one felt too significant to keep to myself.

When your “MCP moment” arrives, you'll understand exactly why.

—Dharmesh (@dharmesh)