What Is MCP and Why You Should Care
MCP evolves beyond traditional APIs to help agentic AI gain access to data
MCP seeks to simplify how LLMs ask for and receive data.
Almost two years to the day after the release of ChatGPT, on November 25, 2024, Anthropic released the open-source Model Context Protocol (MCP). While garnering much less attention than the chat tool, MCP looks like it will have an impact, if not more so, on the AI revolution. Why?
In layman’s terms, MCP makes it much easier for AI agents to access systems and data. This allows an LLM querying against tasks to move beyond clunky “1.0” automated data access via traditional APIs to a much more nimble method that matches the immediacy of the task. MCP is becoming the “data language” of AIs.
Access to domain and organization-specific has been the Achilles heel of meaningful corporate adoption of AI, particularly for small and medium enterprises. Large public enterprises have certain advantages regarding AI based on resources. Smaller sub-billion-dollar companies and nonprofits lack the infrastructure to nimbly access development resources to clean and connect data sources to serve AI deployments. In many cases, even data governance is an aspirational concept.
When you cannot access data easily, AIs cannot function optimally. In the words of Anthropic, “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 custom implementation, making truly connected systems difficult to scale.”
In the best cases, MCP access facilitates easy “if this, then that” connectivity between AI chatbot services like Claude and ChatGPT organizations. Zapier has already adopted the MCP protocol and boasts that its service allows AI agents to connect with more than 7,000 applications. The chief selling point for Zapier? No APIs are needed.
Widespread Industry Adoption of MCP
Hash Studioz explains MCP versus API.
While still relatively new, widespread adoption of the MCP protocol is unfolding before our eyes. Anthropic’s chief rival OpenAI embraced MCP at the end of March and made it a core element of its Responses API for agentic AI just last week.
In addition to Zapier and OpenAI, companies like Hubspot, Salesforce, Cloudflare, PayPal, Wix, Microsoft, and Amazon Web Services have embraced MCP. Rival standards from Google and Cisco have struggled to achieve the same level of adoption. MCP’s momentum is so strong that Google Cloud just announced support for the data transfer protocol despite its rival A2A protocol connectivity efforts.
Because MCP has become the prevalent connector for an AI-fueled Internet, we can expect rapid adoption of platforms and SaaS providers. This will provide customers with much faster access to their domain-specific data.
MCP is not perfect yet, and rapid adoption will likely force evolutions in the protocol, at least in the way it is set up on internal servers, Internet platforms, and SaaS providers. Perhaps the biggest concern is the protocol’s weak security measures. Easy-to-access MCP creates numerous risks for an organization, which, as the implementer, bears the onus of provisioning security to protect its systems from hackers and other black hat operations. Facilitating ease of use and access to data can be pretty inviting for bad actors.
Still, given widespread MCP adoption, the barriers to more relevant and contextual AI for corporations and consumers will become much faster. In many ways, MCP becomes a lubricant for more relevant, impactful AI models.
How It Works: The Nerdy Version of MCP
As a non-developer business person, this is my understanding of MCP (grain of salt applied). The protocol does more than just pull data, going beyond to facilitate access and using AI to analyze data flows. MCP provides a standardized way for AI models to access external tools and data sources through defined interfaces.
noted MCP allows the AI to self-discover on a dynamic basis in real time. The connectivity facilitates an LLM’s ability to better act like an agent, learn, and choose the best outcomes based on its task goal. See his excellent chart below.
According to Andreessen Horowitz: “MCP is an open protocol that allows systems to provide context to AI models in a manner that’s generalizable across integrations. The protocol defines how the AI model can call external tools, fetch data, and interact with services… Based on the context, AI agents can decide which tools to use, in what order, and how to chain them together to accomplish a task.”
MCP requires developers to set up protocol-specific servers to provision access to platforms. Thus, major platforms and environments are essential in provisioning MCP access. Once server access is provisioned through authenticated accounts, MCP effectively serves as a pipeline allowing the AI and data source system to interact.
At that point, it’s up to the organization or user to provision the agent or LLM access to the data source(s). Currently, the primary users of MCP are developers, which makes sense. However, as the systems improve over the ensuing months, it is expected that business users such as myself will be able to facilitate integrations simply using account authentication.
There are a few significant hurdles for MCP or its users to overcome to allow the protocol to become more successful:
Security: MCP facilitates access, but this is bit of a Wild West initiative. MCP implementations can include authentication and authorization mechanisms - the challenge is ensuring proper security practices in implementations. That leaves the implementing organization to create and meter security, which brings additional risks.
UI/UX is built for LLMs, not humans: A pipeline built for AIs has its issues, including uncontrolled risky content and unlimited data.
Web-provisioned data is weak: If a server is not used to facilitate access to a platform, a data connection can occur using MCP over the traditional web, but data access is slow and wonky.
These challenges will be addressed as the major platforms become involved. For example, Okta is working with Google to secure MCP actions in the GCP environment. Cloudflare is working to provide web-based access to content by integrating the new Streamable HTTP transport protocol, a recent update designed to improve web provisioning.
Conclusion
MCP has made access to data much easier for agentic AI and simple LLM querying. While technical challenges must be addressed to make MCP workable on a macro scale, its ability to connect AI models with existing platforms in a simple fashion makes it irresistible.
A world of data that was previously problematic to access will now become available. This makes AI much more useful to a wider group of organizations, companies, and nonprofits that simply could not afford the developer resources necessary to clean and integrate their data.
And because of that, I think more and more developers are going to latch on to it to hasten agentic AI.
Are there players out there now selling MCP’s or is it an open source thing? How are MCP’s different from API’s?
I just tried it and connected a few platforms. I’m not sure of the value yet but will I’m sure.