MCPs—Model Context Protocols—are set to transform AI from passive chatbots into powerful, action-taking agents. But the real story isn’t what MCPs enable—it’s who controls them.
For API-to-MCP Providers, you're right, I think this will be a good opportunity, like nginx for a backend service. I've just open-sourced an MCP-Gateway here https://github.com/mcp-ecosystem/mcp-gateway :)
The control from OpenAI and Claude can be taken away, by hosting Llama and its flavours, and with other models, just like cursor or github copilot is using many models and the user has the choice. The direction is pretty clear, the company which can add as many meaningful MCPs will win this.
I agree with the developer experience issues you mention here.
You also hit on an important point subtly - IMO the strongest value prop for MCPs seem to be distribution. i.e., a developer creates an MCP and it is "available" to many users via existing clients (Claude, Cursor etc).
This is different from the process of building independent standalone apps - e.g., it is unlikely that the overhead of writing an MCP server provides benefits for the developer shipping a bundled app.
If you missed it there is a ”YOLO” mode in Cursor that will execute MCP requests without a confirmation (do not remember the exact setting - I am on the phone).
Yeah I know of that but it scares me as I think it enables all MCP requests which I don't like. Also not sure if I would turn that on for everyone right now without some vetting?
Sounded like it was a complain that it was not there but I agree enabling it could be a security risk. That said for now at least Cursor will only run configured MCP servers (or?!) so as long as you only run your own or vetted it would be kind of low risk.
I can share the learning one that attaches to Cursor. It is pretty prototype-ish and was completely vibe-coded so not hard to recreate using Claude.
The Prediction market one I can pass along the SSE link if people want, but the code is embedded on the betsee.xyz site using MCP libraries and the link (in Elixir)
> Installation, meanwhile, will resemble the mobile app store model. ... Companies will compete—and likely pay significant sums—for premium visibility in these ecosystems, turning MCP directories into high-stakes distribution platforms.
>
> Users will be able to install MCPs—or “chat apps”—from large, curated directories.
I'd predict that a vast swath of users won't want or need to install MCPs — instead they'll use the default AI experience. The provider of that experience will decide which MCPs to install. The quality and power and usefulness of that AI experience will be determined by how good those MCPs are. But it'll be a B2B interaction, not B2C.
Of course there will be power users who prefer an AI experience where they get to choose amongst various MCP options. Just like today there are people who use DuckDuckGo or Linux. But the majority will be happy to effectively delegate those choices to someone who likes thinking about such things.
I also wrote in that post: "While installation is technically possible, most users won’t bother to choose their own tools. Instead, they’ll rely on the defaults provided by the client (like ChatGPT). These defaults won’t be arbitrary—they’ll be the result of lucrative partnerships." And DuckDuckGo was an example I had put in earlier and then removed to shorten it.
I guess our difference in not so much in what will happen but in the relative emphasis and importance of things. We agree that most users will rely on defaults provided by the client — which makes me lose interest in the New Ecosystem Opportunities listed in the article, and become much more interested in how the upcoming mass adoption of AI clients will play out. What are the killer apps, what makes for stickiness, etc.
Also you predict that "these open clients will remain... influential" and I would love for that to be true but I'm having a hard time steelmanning it... the influence of DuckDuckGo and Linux has seemed pretty subtle to me. Certainly open clients will be the Petri dish for awesome new possible things, incubation of cool new kinds of agents and agent services. So, like 4chan, maybe something interesting will emerge from the ooze and get picked up by the majors.
That is fair. I believe that if this world plays out, some MCP server categories will still be big businesses.
1) MCP-First companies. I believe there are going to be MCP-Server-first companies that do something with MCPs that is not doable via mobile app or web. Uber for mobile is a great example. They will use data not available to the MCP Clients (or is hard for them) and will have network effects or other data moats making them monopolies. Those types of servers will have power and big businesses. As an example, imagine if X did not have Grok - there would be a premium to have an LLM be able to search X in real-time.. My hunch is that there will be a few other types of these businesses, including a few startups, just like there have been with every major technical transformation. We just don't know who they are yet.
2) API-to-MCP. I agree that the % of consumer "open-source clients" will be tiny (like the 3-5% marketshare range max). But... I believe a lot of enterprise software will be built on top of open-source MCP clients for privacy and security reasons. Those servers will require multiple MCP points (think of MCPs for each part of their company software stack, including internal components). If that happens, there will be successful businesses that make it easy to go API-MPC. And if we need API to MPC, we will need analytics and security layers (hence Cloudflare - which I found out already has built this haha :)).
3) For affiliate MCPs servers - I was hesitant to put that in - but I bet a few companies come along and strike deals with OpenAI and Claude and just become entrenched as their shopping provider. Also, I could imagine some vertical MCPs (like travel agents) using an affiliate shopping partner (get your bathing suits now that you are going to Maui!)
Still a lot of opportunities for servers IF MCPs take off - much like many big businesses were built off of the iOS and Android and even Salesforce app stores - even though iOS and Android dominate. The real question is IF they take off or not.
I will add thinking of what clients to build is also a great idea!
Cloudflare is already working on it right now: https://developers.cloudflare.com/agents/guides/remote-mcp-server/
For API-to-MCP Providers, you're right, I think this will be a good opportunity, like nginx for a backend service. I've just open-sourced an MCP-Gateway here https://github.com/mcp-ecosystem/mcp-gateway :)
The control from OpenAI and Claude can be taken away, by hosting Llama and its flavours, and with other models, just like cursor or github copilot is using many models and the user has the choice. The direction is pretty clear, the company which can add as many meaningful MCPs will win this.
I agree with the developer experience issues you mention here.
You also hit on an important point subtly - IMO the strongest value prop for MCPs seem to be distribution. i.e., a developer creates an MCP and it is "available" to many users via existing clients (Claude, Cursor etc).
This is different from the process of building independent standalone apps - e.g., it is unlikely that the overhead of writing an MCP server provides benefits for the developer shipping a bundled app.
https://newsletter.victordibia.com/p/no-mcps-have-not-won-yet
If you missed it there is a ”YOLO” mode in Cursor that will execute MCP requests without a confirmation (do not remember the exact setting - I am on the phone).
Yeah I know of that but it scares me as I think it enables all MCP requests which I don't like. Also not sure if I would turn that on for everyone right now without some vetting?
Sounded like it was a complain that it was not there but I agree enabling it could be a security risk. That said for now at least Cursor will only run configured MCP servers (or?!) so as long as you only run your own or vetted it would be kind of low risk.
Can you share the code of the MCPs you built?
I can share the learning one that attaches to Cursor. It is pretty prototype-ish and was completely vibe-coded so not hard to recreate using Claude.
The Prediction market one I can pass along the SSE link if people want, but the code is embedded on the betsee.xyz site using MCP libraries and the link (in Elixir)
My prediction differs from yours at this point:
> Installation, meanwhile, will resemble the mobile app store model. ... Companies will compete—and likely pay significant sums—for premium visibility in these ecosystems, turning MCP directories into high-stakes distribution platforms.
>
> Users will be able to install MCPs—or “chat apps”—from large, curated directories.
I'd predict that a vast swath of users won't want or need to install MCPs — instead they'll use the default AI experience. The provider of that experience will decide which MCPs to install. The quality and power and usefulness of that AI experience will be determined by how good those MCPs are. But it'll be a B2B interaction, not B2C.
Of course there will be power users who prefer an AI experience where they get to choose amongst various MCP options. Just like today there are people who use DuckDuckGo or Linux. But the majority will be happy to effectively delegate those choices to someone who likes thinking about such things.
I think we are in agreement?
I also wrote in that post: "While installation is technically possible, most users won’t bother to choose their own tools. Instead, they’ll rely on the defaults provided by the client (like ChatGPT). These defaults won’t be arbitrary—they’ll be the result of lucrative partnerships." And DuckDuckGo was an example I had put in earlier and then removed to shorten it.
I guess our difference in not so much in what will happen but in the relative emphasis and importance of things. We agree that most users will rely on defaults provided by the client — which makes me lose interest in the New Ecosystem Opportunities listed in the article, and become much more interested in how the upcoming mass adoption of AI clients will play out. What are the killer apps, what makes for stickiness, etc.
Also you predict that "these open clients will remain... influential" and I would love for that to be true but I'm having a hard time steelmanning it... the influence of DuckDuckGo and Linux has seemed pretty subtle to me. Certainly open clients will be the Petri dish for awesome new possible things, incubation of cool new kinds of agents and agent services. So, like 4chan, maybe something interesting will emerge from the ooze and get picked up by the majors.
That is fair. I believe that if this world plays out, some MCP server categories will still be big businesses.
1) MCP-First companies. I believe there are going to be MCP-Server-first companies that do something with MCPs that is not doable via mobile app or web. Uber for mobile is a great example. They will use data not available to the MCP Clients (or is hard for them) and will have network effects or other data moats making them monopolies. Those types of servers will have power and big businesses. As an example, imagine if X did not have Grok - there would be a premium to have an LLM be able to search X in real-time.. My hunch is that there will be a few other types of these businesses, including a few startups, just like there have been with every major technical transformation. We just don't know who they are yet.
2) API-to-MCP. I agree that the % of consumer "open-source clients" will be tiny (like the 3-5% marketshare range max). But... I believe a lot of enterprise software will be built on top of open-source MCP clients for privacy and security reasons. Those servers will require multiple MCP points (think of MCPs for each part of their company software stack, including internal components). If that happens, there will be successful businesses that make it easy to go API-MPC. And if we need API to MPC, we will need analytics and security layers (hence Cloudflare - which I found out already has built this haha :)).
3) For affiliate MCPs servers - I was hesitant to put that in - but I bet a few companies come along and strike deals with OpenAI and Claude and just become entrenched as their shopping provider. Also, I could imagine some vertical MCPs (like travel agents) using an affiliate shopping partner (get your bathing suits now that you are going to Maui!)
Still a lot of opportunities for servers IF MCPs take off - much like many big businesses were built off of the iOS and Android and even Salesforce app stores - even though iOS and Android dominate. The real question is IF they take off or not.
I will add thinking of what clients to build is also a great idea!