MCP and the Innovation Paradox: Why Open Standards Will Save AI itself

The larger models aren’t driving the next wave of AI innovation. The real confusion is quiet: standardization.

The Model Context Protocol (MCP), launched by humanity in November 2024, standardizes the way AI applications interact with the world beyond training data. Just as HTTP and REST standardize how web applications connect to services, MCP standardize how AI models connect to tools.

You’ve probably read dozens of articles explaining what an MCP is. But most mistakes are boring and powerful part: MCP is the norm. Standards aren’t just about organizing technology. They create a growth flywheel. Adopt them early and get on the waves. Ignore them and you’re late. This article explains why MCP is important, the challenges it introduces, and how it is already restructuring the ecosystem.

How MCP moves from chaos to context

Meet Lily, Product Manager at Cloud Infrastructure Company. She juggles projects across half a dozen tools, including Jira, Figma, Github, Slack, Gmail, and Confluence. Like many, she is owned to updates.

By 2024, Lily saw how good the Large-scale Language Model (LLM) is in information integration. She found the opportunity. If she could send all the team’s tools to the model, she could automate updates, draft communications, and answer questions in demand. However, all models had custom methods to connect to services. Each integration has led her to deepen into one vendor’s platform. When she needed to draw transcripts from the gong, it meant building yet another bespoke connection, making it even more difficult to switch to a better LLM later.

Humanity then launched the MCP: an open protocol to standardize the way context flows to LLM. MCP received immediate support Openai, aws, azure, Microsoft Copilot Studio And immediately, Google. Official SDK is available Python, Typescript, Java, C#, rust, Kotrin and Fast. Community SDK go And then others followed. The adoption was quick.

Today, Lily runs everything through Claude and is connected to the work app via a local MCP server. The status report drafts itself. Leadership renewal is quick. When new models appear, she can exchange them without losing integration. When she writes the code on the side, she uses the cursor using the same model of the MCP server that she does with Claude. Her IDE already understands the products she is building. The MCP made this simple.

Standard power and meaning

Lily’s story shows a simple truth. No one likes to use fragmented tools. No user likes to be locked by vendors. Also, no company wants to rewrite integrations every time they change their model. You want the freedom to use the best tools. MCP will deliver.

Now, there is a standard.

First, SaaS providers without strong public APIs are vulnerable to obsolescence. MCP tools rely on these APIs, and customers request support for AI applications. There is no excuse as de facto standards are emerging.

Second, the AI ​​application development cycle is about to dramatically speed up. Developers no longer need to write custom code to test simple AI applications. Instead, you can integrate your MCP server with easily available MCP clients such as Claude Desktop, Cursor, and Windsurf.

Third, switching costs have collapsed. Because integration is detached from a particular model, organizations can migrate from Claude to Openai or Gemini or Blend models without rebuilding their infrastructure. Future LLM providers will benefit from the existing ecosystem around MCP and be able to focus on improving price performance.

Navigate issues with MCP

All standards introduce new friction points or do not resolve existing friction points. MCP is no exception.

Trust is important: Numerous MCP registries have emerged, providing thousands of servers maintained in the community. However, if you do not control the server or trust the parties that do it, you risk revealing secrets to unknown third parties. If you are a SaaS company, provide an official server. If you are a developer, look for the official server.

The quality is variable: APIs have evolved, and insufficiently maintained MCP servers can easily fall out of sync. LLMS relies on high-quality metadata to determine which tools to use. There is no authoritative MCP registry yet to reinforce the need for official servers from trustworthy political parties. For SaaS Company, keep your servers as your API evolves. If you are a developer, look for the official server.

Big MCP servers increase cost and utility reduction: Bundling too many tools into a single server increases costs through token consumption, overwhelming models with too many choices. LLM can easily be confused when you have access to too many tools. It’s the worst of both worlds. Smaller, task-centric servers are important. Keep this in mind when building and distributing servers.

The challenges of approval and identity persist: These issues existed before the MCP, but still exists in the MCP. Imagine that Lily gave Claude the ability to send emails and gave intentional instructions such as “Speed ​​Chris quickly to send status updates.” Instead of sending an email to your boss, Chris, LLM sends an email to everyone named Chris to their contact list so that Chris receives the message. Humans need to stay in the loop for high judgment actions.

Looking ahead

MCP is not a hype, it is a fundamental change in the infrastructure of AI applications.

And like all the well-employed standards before that, MCP is creating self-enhancing flywheels. All new servers, all new integrations, all new applications are tearing the momentum.

New tools, platforms, and registries have already emerged to simplify building, testing, deployment and discovery of MCP servers. As the ecosystem evolves, AI applications provide simple interfaces to connect to new features. Teams accepting protocols will ship products faster with a better integrated story. Companies that provide public APIs and official MCP servers can become part of the integrated story. Slow recruits must fight for relevance.

Noah Schwartz is the head of the product Postman.