Hey, I’m Colin! I help PMs and business leaders improve their technical skills through real-world case studies. Subscribe to paid to get access to the AI Product Circle community on Slack, my full Tech Foundations for PMs course (84 lessons), and a prompt library for common PM tasks.
If you’re not paying close attention, you may have missed a critical update to ChatGPT and Claude that enables a brand new distribution channel: Integrations.
OpenAI announced support for remote MCP integrations on June 4th. Not to be outdone, Claude announced the same on June 18th.
In this post, I’ll describe what MCP is, it’s current state of support, and how you can start to think about incorporating MCP support into your product. I’ll also share a vision of the future, including parallels to the early days of the app store and the potential for tool sprawl within ChatGPT and Claude.
Let’s dive in!
What is MCP?
MCP (Model Context Protocol) is a standardized protocol for communication between agents and external servers. The role of MCP is to define the common ideas and requests that an agent can use. These include:
Tools: actions the agent can take
Prompts: templated messages
Resources: files and other static data
Today, tools are the most commonly implemented feature of MCP. Think of asking ChatGPT to pull your financial projections from Google Sheets and forecast three new scenarios. The ability for agents like ChatGPT and Claude to interact with your data where it already lives opens up a whole new set of possibilities.
Here’s how it works:
A user makes a query inside of ChatGPT or another agent.
The agent checks the MCP servers it's connected to and determines if any tools would be relevant in resolving the request.
The agent constructs the request to the MCP server by calling the tool and adding any necessary parameters. The MCP server can either resolve the request itself, or proxy the request back to a regular web server to interact with your existing API.
The chain gets resolved, with a response back from the app server to the MCP server, MCP to the agent, and finally the agent uses the returned information to write a response to the user.
Here’s a quick example of running this through in Claude with an MCP connection to Gmail.
In this example, I provide Claude a query. It uses this information to send a request to a Gmail MCP server with a tool call for searching messages. This likely hits an existing Gmail API, returns the content, and then Claude summarizes the information for me.
For most of the past year, MCP servers were run on local machines and products like ChatGPT did not support MCP connections.
In June, that changed with the addition of remote MCP servers. Remote MCP is simply an MCP server that’s hosted on the internet, rather than on your computer. The main difference is the setup; remote MCP allows regular users to paste a link and connect their favourite tools. It also allows you to setup authentication, ensuring user data stays secure.
MCP is quickly becoming much more popular. This image shows the growth of popular libraries developers use for building MCP servers, a proxy for the growth rate of MCP servers being built.
As a comparison, I’ve included the growth of Express JS, one of the most popular frameworks for building regular web servers.
Current state of remote MCP support
As exciting as remote MCP support seems, it's still early days. Today, Claude allows any Pro user to add an external URL as an MCP server. The server must support OAuth, even if it returns a public access token.
Here’s what the configuration looks like:
ChatGPT makes MCP connections a little more challenging. First, the UI to access custom connectors is buried behind first party connections. Second, these connections only work in Deep Research mode, meaning most users would never see them used. Finally, this setting is currently only accessible to those on a Team plan.
This is unfortunate, as ChatGPT is the dominant consumer application and exposing your product through it would be a meaningful way to add new value and acquire new customers.
Here’s the process to add a connection from a user’s perspective:
What MCP means for you
Brian Balfour recently shared a great post on the next great distribution shift. He believes that these connections within AI agents represent the next wave of distribution, after which OpenAI and Anthropic will commoditize the most popular use cases. Here’s what he has to say:
“The stampede is coming. If I’m right, in six months, it’s likely every SaaS product and consumer application will be rushing to complete a ChatGPT integration. In twelve months, users will expect it. In eighteen months, the platform taxes will arrive. In twenty-four months, the graveyard will be full.”
I share a similar perspective to Brian, but I think there will be a greater variety of outcomes based on the types of applications built. I think we’ll end up with roughly four types of MCP applications:
Small utilities & apps
Mega corp integrations
Data-driven apps
New and unknown
Small utilities & apps
Think early iPhone apps, like Shazam. These will likely be the first to get cannibalized by the platforms themselves. Why maintain a separate tip calculator app when ChatGPT can do math? The barrier to entry is low, but so is the defensibility.
Mega corp integrations
Companies like Salesforce, Microsoft, and Google will build comprehensive MCP integrations that become table stakes. These integrations will be strategic partnerships rather than acquisition targets, as the platforms need these relationships more than they need to own the functionality.
Data-driven apps
Applications that primarily provide access to proprietary data sets or specialized workflows. Financial data providers like Bloomberg, specialized databases like Crunchbase, or industry-specific tools will maintain value through their unique data and domain expertise.
New and unknown
The most interesting category will be entirely new types of applications that wouldn't make sense without AI agents. These might be workflow orchestrators, multi-app data synthesizers, or entirely new product categories we haven't imagined yet.
The winners in this space will likely be companies that can aggregate multiple data sources or provide comprehensive workflows rather than point solutions. Just as we saw consolidation in the mobile app ecosystem, we'll probably see similar patterns emerge with MCP integrations.
If you're considering building an MCP integration, here are some questions to guide your strategy:
Is your product primarily about data access or workflow? Data-heavy products are more likely to maintain long-term value through MCP than simple utility functions.
Can you create network effects? Integrations that become more valuable as more people use them (like collaboration tools) have better defensibility.
How easily can your core functionality be replicated? If your main value prop can be built into the AI platform directly, consider focusing on more complex, differentiated features.
What's your relationship with the platform? Large companies should pursue strategic partnerships, while smaller companies might focus on becoming indispensable before getting acquired or copied.
MCP represents a fundamental shift in how software is distributed and consumed. Just as the app store changed mobile development and the web transformed desktop software, MCP could reshape how we interact with digital tools entirely.