Adoption of the Model Context Protocol (MCP) has exploded since Anthropic's initial release just seven months ago. As organizations rush to integrate MCP into their AI workflows and overall product offerings, understanding the best practices for implementation becomes crucial.
API strategy consultant Kevin Swiber, who has spent 15 years in the API space working with companies like Postman and advising the OpenAPI initiative, shares valuable insights on how to approach MCP design effectively why it's definitely not going to be the API killer.
If you'd prefer to watch Martyn & Kevin's conversation, you can below!
Understanding MCP's Role in the Technology Stack#
One of the biggest misconceptions about MCP is that it will replace traditional APIs. This assumption follows a familiar pattern we've seen with previous technologies. When GraphQL emerged, many predicted it would kill RESTful APIs, yet both technologies coexist successfully today.
MCP represents a new interface layer rather than a replacement technology. It sits alongside existing RESTful APIs, GraphQL endpoints, and other services, providing a bridge between AI chat interfaces and backend systems. Think of it as another abstraction layer, similar to how API gateways have been used to provide modern interfaces for legacy SOAP services while keeping the underlying systems intact.
The API Multiplication Effect#
Rather than reducing API usage, MCP implementations often become significant API consumers. A single MCP operation frequently requires multiple backend API calls to accomplish its task. This means that instead of killing the API economy, MCP could actually drive increased API adoption and usage.
This presents an opportunity for existing API developers who may have wondered how to participate in the AI revolution without becoming machine learning engineers. MCP provides a pathway for traditional developers to leverage their existing API knowledge and contribute to AI-powered experiences.
What to Consider When Building for MCP#
Context Window Management#
Unlike traditional API design where response size primarily affected network performance and user experience, MCP introduces the constraint of context window limitations. Large responses can consume valuable token space that could be better used for reasoning and task completion. This requires rethinking how we structure and size our responses.
Tool Selection Constraints#
Large Language Models struggle with decision-making when presented with too many options. If your API has 100 different operations and you expose each one as an MCP tool, the LLM will have difficulty selecting the appropriate tool for a given task. This necessitates careful curation and grouping of functionality.
The design challenge becomes creating higher-level, more semantic operations rather than exposing every granular API endpoint.
Think in terms of user intentions rather than system capabilities.
MCP and the Developer Experience Evolution#
New Entry Points for Learning#
The traditional developer journey that typically begins with documentation is changing. Developers are increasingly starting their learning process through chat interfaces like ChatGPT or Claude rather than visiting company documentation sites. This shift requires organizations to optimize their content for AI consumption and consider how their information will be discovered and presented through these new channels.
Simplified Dev Tool Onboarding Through AI#
It's already possible to get recommendations on developer tools to solve specific tasks from AI today, and we're approaching a future where the entire developer onboarding process (from service discovery to account creation to API key generation) could happen entirely within a chat interface. This seamless experience would eliminate the need for developers to navigate multiple websites and forms, potentially accelerating adoption significantly (if done well).
Security as a Foundational Requirement#
Security concerns around MCP are substantial, particularly with remote MCP servers. Organizations are approaching MCP adoption through two primary lenses: risk mitigation and value creation. Security teams are actively working to establish best practices, and some previously underutilized authorization standards are gaining renewed attention as organizations seek to implement MCP securely.
The current state of MCP configuration (primarily through JSON file editing for local usage) makes it primarily accessible to technical users rather than business stakeholders.
This technical barrier actually provides some inherent security benefits. However, remote MCP server adoption has landed in multiple major players UIs already so it's doubtful that's going to last too long.
The OpenAPI Bridge#
OpenAPI specifications serve as an excellent gateway for newcomers to MCP development. These human-readable documents can function similarly to the "view source" feature that helped early web developers learn HTML. With modern AI tools, developers can generate high-quality OpenAPI documents from simple prompts, then use these specifications to create MCP servers quickly.
This approach provides a familiar foundation for developers while opening pathways to more advanced concepts like workflow orchestration tools that better match MCP's interaction patterns.
MCP Won't Kill APIs#
MCP represents more than just a new protocol. It's democratizing access to AI development for a broader range of developers. The enthusiasm around MCP echoes the early excitement of web development in the mid-1990s, suggesting we're at the beginning of a significant shift in how developers interact with AI systems.
The rapid pace of innovation in this space means that best practices are still evolving. Organizations should focus on experimentation while maintaining security consciousness, understanding that the patterns and practices we establish now will influence how this technology develops.
For API companies and developers, the message is clear: MCP isn't a threat to existing API investments. It's an opportunity to extend their reach into the rapidly growing world of AI-powered applications. The key is approaching MCP design with intention, considering the unique constraints and opportunities this new interaction model provides.