MCP and the innovation paradox: Why open standards will save AI from itself

Larger fashions aren’t driving the following wave of AI innovation. The true disruption is quieter: Standardization.

Launched by Anthropic in November 2024, the Mannequin Context Protocol (MCP) standardizes how AI functions work together with the world past their coaching information. Very like HTTP and REST standardized how internet functions connect with companies, MCP standardizes how AI fashions connect with instruments.

You’ve in all probability learn a dozen articles explaining what MCP is. However what most miss is the boring — and highly effective — half: MCP is a regular. Requirements don’t simply set up know-how; they create progress flywheels. Undertake them early, and also you trip the wave. Ignore them, and also you fall behind. This text explains why MCP issues now, what challenges it introduces, and the way it’s already reshaping the ecosystem.

How MCP strikes us from chaos to context

Meet Lily, a product supervisor at a cloud infrastructure firm. She juggles initiatives throughout half a dozen instruments like Jira, Figma, GitHub, Slack, Gmail and Confluence. Like many, she’s drowning in updates.

By 2024, Lily noticed how good massive language fashions (LLMs) had change into at synthesizing data. She noticed a possibility: If she may feed all her workforce’s instruments right into a mannequin, she may automate updates, draft communications and reply questions on demand. However each mannequin had its customized approach of connecting to companies. Every integration pulled her deeper right into a single vendor’s platform. When she wanted to drag in transcripts from Gong, it meant constructing one more bespoke connection, making it even more durable to modify to a greater LLM later.

Then Anthropic launched MCP: An open protocol for standardizing how context flows to LLMs. MCP shortly picked up backing from OpenAI, AWS, Azure, Microsoft Copilot Studio and, quickly, Google. Official SDKs can be found for Python, TypeScript, Java, C#, Rust, Kotlin and Swift. Neighborhood SDKs for Go and others adopted. Adoption was swift.

In the present day, Lily runs every part via Claude, related to her work apps by way of an area MCP server. Standing stories draft themselves. Management updates are one immediate away. As new fashions emerge, she will be able to swap them in with out shedding any of her integrations. When she writes code on the facet, she makes use of Cursor with a mannequin from OpenAI and the identical MCP server as she does in Claude. Her IDE already understands the product she’s constructing. MCP made this straightforward. 

The ability and implications of a regular

Lily’s story reveals a easy fact: No one likes utilizing fragmented instruments. No person likes being locked into distributors. And no firm needs to rewrite integrations each time they alter fashions. You need freedom to make use of the perfect instruments. MCP delivers.

Now, with requirements come implications.

First, SaaS suppliers with out sturdy public APIs are weak to obsolescence. MCP instruments rely upon these APIs, and clients will demand assist for his or her AI functions. With a de facto commonplace rising, there are not any excuses.

Second, AI software improvement cycles are about to hurry up dramatically. Builders not have to write down customized code to check easy AI functions. As an alternative, they’ll combine MCP servers with available MCP purchasers, similar to Claude Desktop, Cursor and Windsurf.

Third, switching prices are collapsing. Since integrations are decoupled from particular fashions, organizations can migrate from Claude to OpenAI to Gemini — or mix fashions — with out rebuilding infrastructure. Future LLM suppliers will profit from an current ecosystem round MCP, permitting them to concentrate on higher worth efficiency.

Navigating challenges with MCP

Each commonplace introduces new friction factors or leaves current friction factors unsolved. MCP isn’t any exception. 

Belief is crucial: Dozens of MCP registries have appeared, providing hundreds of community-maintained servers. However if you happen to don’t management the server — or belief the occasion that does — you danger leaking secrets and techniques to an unknown third occasion. In case you’re a SaaS firm, present official servers. In case you’re a developer, search official servers.

High quality is variable: APIs evolve, and poorly maintained MCP servers can simply fall out of sync. LLMs depend on high-quality metadata to find out which instruments to make use of. No authoritative MCP registry exists but, reinforcing the necessity for official servers from trusted events. In case you’re a SaaS firm, keep your servers as your APIs evolve. In case you’re a developer, search official servers.

Massive MCP servers enhance prices and decrease utility: Bundling too many instruments right into a single server will increase prices via token consumption and overwhelms fashions with an excessive amount of selection. LLMs are simply confused if they’ve entry to too many instruments. It’s the worst of each worlds. Smaller, task-focused servers will likely be necessary. Preserve this in thoughts as you construct and distribute servers.

Authorization and Id challenges persist: These issues existed earlier than MCP, and so they nonetheless exist with MCP. Think about Lily gave Claude the flexibility to ship emails, and gave well-intentioned directions similar to: “Shortly ship Chris a standing replace.” As an alternative of emailing her boss, Chris, the LLM emails everybody named Chris in her contact record to ensure Chris will get the message. People might want to stay within the loop for high-judgment actions.

Wanting forward

MCP isn’t hype — it’s a basic shift in infrastructure for AI functions.

And, similar to each well-adopted commonplace earlier than it, MCP is making a self-reinforcing flywheel: Each new server, each new integration, each new software compounds the momentum.

New instruments, platforms and registries are already rising to simplify constructing, testing, deploying and discovering MCP servers. Because the ecosystem evolves, AI functions will supply easy interfaces to plug into new capabilities. Groups that embrace the protocol will ship merchandise quicker with higher integration tales. Corporations providing public APIs and official MCP servers may be a part of the combination story. Late adopters must struggle for relevance.

Noah Schwartz is head of product for Postman.