
Why MCP Won the Agent Tooling Wars (And How It Will Speed Up Agents)
What this covers
This is an 11-minute solo explainer from The AI Daily Brief, built around OpenAI's announcement that it will support the Model Context Protocol (MCP)) across its products. The host first sets out what MCP is: Anthropic introduced it in November as an open standard for connecting AI assistants to the places data lives, such as content repositories, business tools, and developer environments. The problem it targets is that models stay isolated behind information silos, and each new data source has needed its own custom integration. The host frames MCP as a universal adapter. An agent's request is translated into whatever format a given API) expects, and the response is translated back into one readable form. The example is Cursor, which would otherwise need separate integrations for GitHub, Slack, and the local file system.
The second half turns to the strategic story. On March 11th OpenAI released an Agents SDK and Responses API with proprietary web search and computer use, which many read as a competing direction. Sam Altman then announced MCP support, available at once in the Agents SDK, with the ChatGPT desktop app and Responses API to follow. The host explains this through earlier format wars such as USB versus Lightning and DVD versus Blu-ray. In those, the argument goes, consensus matters more than technical merit, and network effects decide the outcome. A contested claim is that a standard from a large lab beats a startup's because outside companies fear lock-in less.
The host closes on consequences. Agent builders can start from "first base" by plugging into existing MCP servers instead of writing tool integrations. Companies that favor speed over owning infrastructure should bring agentic applications to market sooner. Tools like Lovable and Bolt are widening who can build with code, so non-technical people will increasingly meet MCP too.
OpenAI's adoption of MCP (Model Context Protocol) as an open standard, rather than pursuing proprietary competition, has consolidated agent tooling around a single platform, enabling faster development and broader agent deployment across the industry.
- MCP achieved critical mass through network effects before OpenAI's decision, making adoption the strategic choice over competition
- Open standards in software reduce integration costs and enable all developers to optimize in a single direction, accelerating innovation
- OpenAI prioritizes speed-to-market and agent capability over owning proprietary infrastructure
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The main lesson from earlier format wars in tech (like USB vs Lightning, DVD vs Blu-ray) is that ultimately it doesn't matter what the standard is technically—the important thing is that there's a consensus, because consensus enables everyone to optimize in a single direction.
“One of the big lessons from earlier format wars in tech is that ultimately it doesn't really matter what the standard is. The important thing is that there's a consensus. We now have both OpenAI and Anthropic sending a big signal to every software company in the world. It's time to build an MCP server and let the agents in. You might not notice it happening if you're not in the weeds coding and deploying infrastructure for agents, but looking back in 3 months time, we are likely to see an astonishing proliferation of new agentic features that are available due to adding MCP support. With everyone now building on the same standard, everyone can optimize in a single direction.”
The breadth of people who can create with code is radically expanding as a result of AI agent tools like Lovable and Bolt, and part of this expansion means that non-technical people will increasingly interact with MCP
“Part of the great transformation that is happening with AI and agents right now is that the breadth of people who can create with code is radically expanding.”
A standard from a large AI lab is more likely to succeed than a standard from a smaller company, regardless of which standard is technically superior, because external companies will adopt a standard from a company that is perceived as too big to lock them in over one from a startup with financial incentives to lock them into their platform
“Touching on why this didn't come from one of these smaller companies that were first to market like Composio, Leighton Space wrote, 'This one is perhaps the most depressing for idealists who want the best idea to win. A standard from a big lab is very simply more likely to succeed than a standard from anyone else. There's nothing fair about this. If the financial future of your startup incentivizes you to lock me into your standard, I'm not adopting it. If the standard backer seems too big to really care about locking you into the standard, then I will adopt it.'”
MCP abstracts away API differences through a universal interpretation layer in the middle, making everything interoperable on the same standard, whereas OpenAI's approach used distinct API calls for different tasks (e.g., a tool use call would look different from a database query call)
“The article dug deep into the technical reasons why MCP was an improvement over the way OpenAI were doing things with one very simple reason being that it just made more sense with the way that AI works. OpenAI was using distinct API calls for different tasks. For example, the call in response to a tool that an agent wants to use would look very different to querying and receiving data from a database. whereas MCP abstracted that all away, adding a universal interpretation layer in the middle, so everything is interoperable on the same standard.”
MCP had both a big backer (Anthropic) and was open standard, whereas OpenAI's solution was functional but locked down within the company, requiring external companies to work with OpenAI on integration and making additions relatively slow
“Leon Space noted that MCP had a combination of not only a big backer, but also an open standard. OpenAI's solution was functional, but was locked down within the company. To get access, tooling and data, companies needed to work with OpenAI on integration. It also meant that additions were relatively slow. In MCP, Anthropic just proposed the open standard and let everyone add themselves.”
OpenAI released a major agentic tooling update on March 11th consisting of an Agents SDK and Responses API with proprietary web search and computer use features, which many initially interpreted as OpenAI building in their own competing direction rather than adopting MCP support
“2 weeks ago, however, on March 11th, OpenAI released their big aentic tooling update, and many thought we were seeing the beginning of a new agent tooling wars. OpenAI's release included a software development kit called agents SDK, as well as a standardized tool use access point called Responses API. These new features allowed agent builders to tap into OpenAI's web search and computer use features. Many people thought this was OpenAI trying to build in their own direction. Basically, instead of just adding MCP support to their models, they had built a set of proprietary tool integrations.”
Anthropic announced the Model Context Protocol in November as a new standard for connecting AI assistance to the systems where data lives, including content repositories, business tools, and developer environments, explicitly framing the problem as AI models being constrained by isolation from data trapped behind information silos and legacy systems
“Back in November, Anthropic announced the model context protocol or MCP. They called it a new standard for connecting AI assistance to the systems where data lives, including content repositories, business tools, and developer environments. MCP was trying to solve a very clear need. Anthropic wrote, 'Even the most sophisticated models are constrained by their isolation from data, trapped behind information silos and legacy systems. Every new data source requires its own custom implementation, making truly connected systems difficult to scale.'”
The adoption of standards in technology ultimately depends more on network effects and breadth of adoption than on which standard is technically superior, following the truism that 'ultimately standard gets adopted isn't necessarily about which standard is best and more about the network effect of just how widely adopted it is'
“And like any network, the more of these servers that came online supporting a broader range of tools, the more that it made sense for developers to just stay inside that ecosystem. MCP, in other words, started to prove the old truism that ultimately standard gets adopted isn't necessarily about which standard is best and more about the network effect of just how widely adopted it is.”
MCP was created to solve the problem that even the most sophisticated AI models are constrained by their isolation from data, trapped behind information silos and legacy systems, with every new data source requiring its own custom implementation, making truly connected systems difficult to scale.
“Even the most sophisticated models are constrained by their isolation from data, trapped behind information silos and legacy systems. Every new data source requires its own custom implementation, making truly connected systems difficult to scale.”
Without access to tools, Large Language Models can't really do anything meaningful in the world other than predict the next token; to move from that to agentic capabilities, they need access to tools and resources.
“Remember that without having access to tools, LLMs can't really do anything meaningful in the world other than predict the next token. To move from that to agentic capabilities, they need access to tools and resources.”
MCP started to prove the truism that the standard that ultimately gets adopted isn't necessarily the best standard, but rather the one with the greatest network effect and widest adoption, as more MCP servers came online supporting a broader range of tools, making it make sense for developers to stay inside that ecosystem.
“MCP, in other words, started to prove the old truism that ultimately standard gets adopted isn't necessarily about which standard is best and more about the network effect of just how widely adopted it is. There are now thousands of MCP servers that allow easy access to basically every major app or tool. And so even before Altman and OpenAI decided to make this announcement, the team at Leighton Space...”
Programmers can spin up MCP servers for specific tools, knowing they won't have to duplicate that work when the next new agent comes along, making MCP effectively a universal adapter for agentic API access that's open for everyone to build on.
“programmers can spin up MCP servers for specific tools, knowing they won't have to duplicate that work when the next new agent comes along. MCP you can kind of think of like a universal adapter for agentic API access that's open for everyone to build on.”
Software standards wars are fundamentally different from hardware standards wars because there is no hardware involved, making it much lower cost to support multiple competing standards simultaneously in software
“But of course, in this case, there's no hardware involved. The standards war is entirely about software integrations, meaning that it's much lower cost to support both.”
The main point of technology standards wars is that it ultimately doesn't matter what the standard is; what matters is that there is consensus, and companies in the industry send a unified signal about which standard to build on
“One of the big lessons from earlier format wars in tech is that ultimately it doesn't really matter what the standard is. The important thing is that there's a consensus. We now have both OpenAI and Anthropic sending a big signal to every software company in the world. It's time to build an MCP server and let the agents in.”
Without access to tools and resources, language models can only predict the next token and cannot perform meaningful actions in the world; agent capabilities require tool access to move beyond token prediction
“Remember that without having access to tools, LLMs can't really do anything meaningful in the world other than predict the next token. To move from that to agentic capabilities, they need access to tools and resources.”
Agent builders will soon be able to plug into the MCP servers they need without having to build tool integrations themselves, meaning they can start on first base and focus only on making their agents work
“Pretty soon, agent builders won't need to build tool integrations at all. They'll just plug into the MCP servers they need and move on. This means that those developers will be starting on first base and can just focus on making their agents work.”
Companies prioritizing speed of development over control of infrastructure means agentic applications will come to market faster
“That means in general that you are going to see more agentic applications come to market faster.”
The breadth of people who can create with code is radically expanding due to the transformation happening with AI and agents.
“Part of the great transformation that is happening with AI and agents right now is that the breadth of people who can create with code is radically expanding.”
Two weeks before March 13th [approximately March 11th], OpenAI released a big agentic tooling update that included a software development kit called agents SDK and a standardized tool use access point called Responses API, allowing agent builders to tap into OpenAI's web search and computer use features.
“2 weeks ago, however, on March 11th, OpenAI released their big aentic tooling update...OpenAI's release included a software development kit called agents SDK, as well as a standardized tool use access point called Responses API. These new features allowed agent builders to tap into OpenAI's web search and computer use features.”
OpenAI has decided to support MCP across its products, with support available today in the agents SDK and support for ChatGPT desktop app and Responses API coming soon, as announced by Sam Altman.
“Yesterday, Sam Alman tweeted, "People love MCP and we are excited to add support across our products available today in the agents SDK and support for chat GBT desktop app and responses API coming soon."”
The Model Context Protocol functions as a universal adapter that converts an agent's request for data into whatever format a specific API requires, then converts the delivered data into a standardized format readable to the agent, eliminating the need for agents to interface with each API's unique format individually
“MCP you can kind of think of like a universal adapter for agentic API access that's open for everyone to build on. An MCP server effectively converts an agent's request for data into whatever format the API is looking for. Then once the data is delivered, it converts it into a standardized format that's readable to the agent.”
Without the Model Context Protocol, an IDE like Cursor has to interact with separate systems like GitHub, Slack, and local file systems each through their own unique API, but with MCP, Cursor only has to interact with MCP through a unified API.
“The first schematic that doesn't involve model context protocol shows cursor having to interact with GitHub, Slack, and a local file system, each through their own unique API. With the model context protocol, on the other hand, MCP is doing the interaction between each of those unique APIs and the end user using cursor only has to interact with MCP through a unified API.”
MCP (Model Context Protocol) is an open standard for connecting AI assistance to the systems where data lives, including content repositories, business tools, and developer environments.
“Back in November, Anthropic announced the model context protocol or MCP. They called it a new standard for connecting AI assistance to the systems where data lives, including content repositories, business tools, and developer environments.”
MCP is an open standard that enables developers to build secure two-way connections between their data sources and AI powered tools through an architecture where developers can either expose their data through MCP servers or build AI applications called MCP clients that connect to these servers.
“Anthropic wrote, MCP is an open standard that enables developers to build secure two-way connections between their data sources and AI powered tools. The architecture is straightforward. Developers can either expose their data through MCP servers or build AI applications called MCP clients that connect to these servers.”
A standard from a big lab is more likely to succeed than a standard from a smaller company (like Composio, which was first to market), not because it's the best idea but because of incentive alignment: if a company's financial future depends on locking you into their standard, developers won't adopt it, but if the standard backer is big enough to not care about locking users in, then developers will adopt it.
“Touching on why this didn't come from one of these smaller companies that were first to market like Composio, Leighton Space wrote, "This one is perhaps the most depressing for idealists who want the best idea to win. A standard from a big lab is very simply more likely to succeed than a standard from anyone else. There's nothing fair about this. If the financial future of your startup incentivizes you to lock me into your standard, I'm not adopting it. If the standard backer seems too big to really care about locking you into the standard, then I will adopt it.”
An MCP server converts an agent's request for data into whatever format the API is looking for, and once the data is delivered, it converts it into a standardized format that's readable to the agent.
“An MCP server effectively converts an agent's request for data into whatever format the API is looking for. Then once the data is delivered, it converts it into a standardized format that's readable to the agent.”
Unlike historical hardware standards wars (USB vs Lightning, DVD vs Blu-ray), the software standards war for agent tooling is much lower cost to support both because there is no hardware involved and the standards war is entirely about software integrations.
“But of course, in this case, there's no hardware involved. The standards war is entirely about software integrations, meaning that it's much lower cost to support both.”
A detailed 100-minute tutorial from the AI Engineer Summit (MCed by the speaker) featuring the Anthropic staff member who designed MCP went viral and explained how to integrate MCP and get best results, which led to an uptick in MCP server development and adoption
“At the end of February, however, a lengthy tutorial seminar from the AI Engineer Summit, the one that I MCed in New York, that featured the anthropic staff member who designed the protocol, started to go viral, or at least as viral as a dense 100minute video aimed at AI engineers can go. It explained exactly how MCP works, how to integrate it, and how to get the best results for agent building. Following this, we started to see an uptick in MCP servers getting built.”
The Model Context Protocol (MCP) has won the agent tooling wars before they even really began, as evidenced by OpenAI's decision to add MCP support across their products after initially appearing to build proprietary alternatives
“With a new announcement from OpenAI, MCP has won the agent tooling wars before they even really began.”
OpenAI has decided that the value of standardizing on MCP in terms of how much faster it allows everyone to move toward really performant agents is worth more than owning that standard exclusively
“TLDDR OpenAI has decided that the value of anointing a standard in terms of how much faster it allows everyone to move towards really performant agents is worth more than owning that standard.”
MCP (Model Context Protocol) has won the agent tooling wars before they even really began because of rapid consolidation around this single open protocol rather than competing proprietary standards.
“With a new announcement from OpenAI, MCP has won the agent tooling wars before they even really began.”
The rapid consolidation around MCP is an indicator that companies in the AI space prioritize moving at the greatest possible speed over trying to battle for ownership of key infrastructure.
“the fact that there has been such quick consolidation around this protocol which is called MCP or model context protocol is an indicator of just how quickly agents are moving and more specifically an indicator that companies in this space would rather move at the greatest possible speed than try to duke it out and try to battle for ownership of key infrastructure.”
Anthropic has become the de facto model and de facto brand for AI engineers, and this is a major reason MCP was adopted and succeeded in competition with OpenAI's approach
“Another reason they argued that MCP was a breakout is the degree to which Anthropic has become the deacto model and the deacto brand for AI engineers. We've talked extensively about how much the developer use case has really become the core of anthropic success and that certainly feels like it's at play here as well.”
Within 3 months of OpenAI's announcement supporting MCP, we will likely see an astonishing proliferation of new agentic features from software companies adding MCP support, because everyone building on the same standard can optimize in a single direction
“You might not notice it happening if you're not in the weeds coding and deploying infrastructure for agents, but looking back in 3 months time, we are likely to see an astonishing proliferation of new agentic features that are available due to adding MCP support. With everyone now building on the same standard, everyone can optimize in a single direction.”
OpenAI appeared to be locking developers into proprietary web search by building their own web search feature (instead of supporting Brave Search via MCP) as part of their March tooling announcement
“To give one example, one of the most popular MCP servers is Brave Search, which allows agents to surf the web. Instead of just letting developers access that through MCP, OpenAI appeared to be locking them into using OpenAI's proprietary web search feature instead.”
The consolidation around MCP will result in more agentic applications coming to market faster.
“That means in general that you are going to see more agentic applications come to market faster.”
The Latent Space team wrote that 'the number one feature of any network is the people already on it,' and accordingly, the power of any new protocol derives from its adoption and ecosystem, with MCP having captured enough critical mass and momentum to already be the presumptive winner of the 2023-2025 agent open standard wars.
“the Latent Space team who overlaps via Swix with the team who runs the AI engineer summit, wrote a blog post called YMCP1. They affirm this same idea that we were just talking about, writing, the number one feature of any network is the people already on it. Accordingly, the power of any new protocol derives from its adoption and ecosystem. And it's fair to say that MCP has captured enough critical mass and momentum right now that it is already the presumptive winner of the 23 to 25 agent open standard wars.”
MCP had a combination of not only a big backer (Anthropic) but also an open standard, whereas OpenAI's solution was functional but locked down within the company, requiring companies to work with OpenAI on integration and making additions relatively slow.
“Leon Space noted that MCP had a combination of not only a big backer, but also an open standard. OpenAI's solution was functional, but was locked down within the company. To get access, tooling and data, companies needed to work with OpenAI on integration. It also meant that additions were relatively slow. In MCP, Anthropic just proposed the open standard and let everyone add themselves.”
Anthropic has become the de facto model and de facto brand for AI engineers, which contributed to MCP's success as the standard.
“Another reason they argued that MCP was a breakout is the degree to which Anthropic has become the deacto model and the deacto brand for AI engineers. We've talked extensively about how much the developer use case has really become the core of anthropic success and that certainly feels like it's at play here as well.”
Pretty soon, agent builders won't need to build tool integrations at all; they'll just plug into the MCP servers they need and move on, meaning developers will start on first base and can just focus on making their agents work.
“Pretty soon, agent builders won't need to build tool integrations at all. They'll just plug into the MCP servers they need and move on. This means that those developers will be starting on first base and can just focus on making their agents work.”
OpenAI has decided that the value of anointing a standard (in terms of how much faster it allows everyone to move towards performant agents) is worth more than owning that standard.
“TLDDR OpenAI has decided that the value of anointing a standard in terms of how much faster it allows everyone to move towards really performant agents is worth more than owning that standard.”
MCP is a technical improvement over OpenAI's approach (using distinct API calls for different tasks) because it abstracted everything away with a universal interpretation layer in the middle, so everything is interoperable on the same standard, whereas OpenAI's approach had different calls for different tasks.
“The article dug deep into the technical reasons why MCP was an improvement over the way OpenAI were doing things with one very simple reason being that it just made more sense with the way that AI works. OpenAI was using distinct API calls for different tasks. For example, the call in response to a tool that an agent wants to use would look very different to querying and receiving data from a database. whereas MCP abstracted that all away, adding a universal interpretation layer in the middle, so everything is interoperable on the same standard.”
There are now thousands of MCP servers that allow easy access to basically every major app or tool
“There are now thousands of MCP servers that allow easy access to basically every major app or tool.”
Within the next 12 calendar months, a fair number of non-technical people will find themselves using tools like Lovable, Bolt, or IDEs like Cursor or Windsurf to interact meaningfully with the Model Context Protocol.
“I think that a fair number of you who are sitting there now as non-technical or not developers will find yourself at some point in the next year, yes I mean the next 12 calendar months using a tool like lovable or bolt or an IDE like cursor or windsurf actually interacting in a meaningful way with the model context protocol.”
In 3 months time (from the episode date), we are likely to see an astonishing proliferation of new agentic features that are available due to adding MCP support, as companies build out their integrations now that consensus exists.
“looking back in 3 months time, we are likely to see an astonishing proliferation of new agentic features that are available due to adding MCP support.”
The adoption of MCP matters to non-developers for two specific reasons: first, it signals that companies prioritize speed of development over battling for infrastructure ownership, leading to faster agentic application deployment; second, non-technical people will directly use MCP-integrated tools like Lovable, Bolt, Cursor, or Windsurf within the next 12 calendar months
“This matters to you as a non-developer for at least two reasons. The first is that the fact that there has been such quick consolidation around this protocol which is called MCP or model context protocol is an indicator of just how quickly agents are moving and more specifically an indicator that companies in this space would rather move at the greatest possible speed than try to duke it out and try to battle for ownership of key infrastructure. That means in general that you are going to see more agentic applications come to market faster.”
One of the most popular MCP servers is Brave Search, which allows agents to surf the web, but instead of letting developers access that through MCP, OpenAI appeared to be locking them into using OpenAI's proprietary web search feature instead.
“To give one example, one of the most popular MCP servers is Brave Search, which allows agents to surf the web. Instead of just letting developers access that through MCP, OpenAI appeared to be locking them into using OpenAI's proprietary web search feature instead.”
MCP had started to pick up steam in the first few months after it was released in November, and it was increasing throughout the end of February and into March.
“MCP had started to pick up steam in the first few months after it was released, and it was increasing throughout the end of February and into March.”
Many observers thought that OpenAI's March 11th announcement was the beginning of a new standards war, comparable to historical tech format wars like USB versus Apple's Lightning connection or DVD versus Blu-ray.
“And so, many thought that we were headed into the beginning of a new standards war. USB versus Apple's Lightning connection, DVD versus Blu-ray.”
A lengthy tutorial/seminar from the AI Engineer Summit (featuring the Anthropic staff member who designed MCP) went viral in late February, explaining how MCP works, how to integrate it, and how to get the best results for agent building, and following this, an uptick in MCP servers getting built was observed.
“a lengthy tutorial seminar from the AI Engineer Summit, the one that I MCed in New York, that featured the anthropic staff member who designed the protocol, started to go viral, or at least as viral as a dense 100minute video aimed at AI engineers can go. It explained exactly how MCP works, how to integrate it, and how to get the best results for agent building. Following this, we started to see an uptick in MCP servers getting built.”
In a few months time, viewers will see concrete benefits from MCP standardization by being able to deploy agents that would not have been possible otherwise
“In just a couple months time, I guarantee you are going to see the benefits of this by being able to deploy agents that would not have been possible otherwise.”
Many people initially thought OpenAI was trying to build in their own direction by creating proprietary tool integrations instead of just adding MCP support to their models.
“Many people thought this was OpenAI trying to build in their own direction. Basically, instead of just adding MCP support to their models, they had built a set of proprietary tool integrations.”
OpenAI's decision to support MCP was not predetermined; MCP was well-received from the beginning, but there was still considerable debate about whether it would become the standard, as evidenced by Langchain's blog post titled 'MCP: Flash in the Pan or Future Standard' published at the end of February.
“Now, it's important to note that this was not predetermined. Like I said, MCP was pretty wellreceived right from the beginning, but there still was a lot of debate. The Langchain blog published something called MCP, flash in the pan or future standard. At the end of February...”
There are now thousands of MCP servers that allow easy access to basically every major app or tool.
“There are now thousands of MCP servers that allow easy access to basically every major app or tool.”
In a couple of months time, users will guarantee to see the benefits of MCP standardization by being able to deploy agents that would not have been possible otherwise.
“In just a couple months time, I guarantee you are going to see the benefits of this by being able to deploy agents that would not have been possible otherwise.”
Latent Space wrote that MCP has already captured enough critical mass and momentum that it is the presumptive winner of the 2023 to 2025 agent open standard wars
“And so even before Altman and OpenAI decided to make this announcement, the team at Leighton Space, who overlaps via Swix with the team who runs the AI engineer summit, wrote a blog post called YMCP1... it is already the presumptive winner of the 23 to 25 agent open standard wars.”
MCP was receiving increasing adoption throughout late February and into March before OpenAI's March 11 announcement
“MCP had started to pick up steam in the first few months after it was released, and it was increasing throughout the end of February and into March.”