From Human-Readable to Machine-Usable: The New API Stack

Sagar Batchu, CEO of Speakeasy, discusses the pivotal shift in API development as AI agents become primary consumers. The conversation covers the rise of the Model Context Protocol (MCP), the challenges in building agent-ready APIs, and how Speakeasy provides a toolchain for creating, managing, and securing MCP servers.

July 17, 2025 4 min read

From Human-Readable to Machine-Usable: The New API Stack — source: https://www.youtube.com/watch?v=zwvEY11aM4s

The Evolving Landscape of API Development

Building high-quality, reliable APIs—on the level of Stripe or Twilio—is an underrated and difficult engineering challenge. Historically, companies relied on a mix of homegrown solutions, “duct tape tooling,” and open-source projects. While larger enterprise solutions like Mulesoft exist, they often impose a top-down, heavyweight process.

The modern challenge is managing API development at scale across multiple teams with competing priorities. This creates a need for a developer-centric toolchain that integrates seamlessly into existing workflows, such as CI/CD and OpenAPI specifications.

The Shift to AI as the Primary API Consumer

A significant driver of change is the emergence of AI and Large Language Models (LLMs) as primary consumers of APIs. For AI to be truly useful, it needs access to real-time, proprietary data and services. This moves beyond initial strategies like fine-tuning or basic Retrieval-Augmented Generation (RAG) to giving models controlled access to APIs for direct interaction and action.

This shift fundamentally expands the total addressable market (TAM) for APIs. Previously limited to developers building integrations, APIs can now be accessed by a much broader, non-technical audience through conversational interfaces, effectively shifting the API’s role “upstream” in the user lifecycle.

Understanding the Model Context Protocol (MCP)

The Model Context Protocol (MCP), championed by Anthropic, is emerging as the leading open standard for enabling “tool use” for LLMs. It provides a standardized specification for how an LLM client can discover and interact with a server that exposes a set of tools (APIs).

Key Differences with OpenAI Function Calling

While OpenAI offers “function calling,” it is a proprietary feature tied to their SDK. MCP’s strength lies in its nature as an open standard. This allows any company to build a compliant server, and any client or LLM to connect to it, fostering a broad and interoperable ecosystem, much like OpenAPI did for REST APIs.

Adoption and Governance

MCP adoption is progressing in waves:

Wave 1 (Experimentation): Initial excitement led to developers running local MCP servers, revealing the complexity and effort required.
Wave 2 (Official Endpoints): Companies began providing official, hosted, first-party MCP servers, offering stability and trust (e.g., HubSpot).
Wave 3 (Ecosystem & Platforms): The current wave involves the rise of platforms, like Speakeasy, that help companies aggregate, create, host, and manage their MCP servers.

Anthropic’s stewardship of the open protocol is viewed as a positive, providing clear direction and quality control, which can accelerate adoption by preventing the fragmentation that sometimes slowed the progress of other open standards.

Building and Managing MCP Servers with Speakeasy

Speakeasy helps companies bridge the gap between their existing APIs and the new world of AI agents. The platform focuses on the entire lifecycle of an MCP server:

Creation: Ingests an existing OpenAPI specification to automatically generate the boilerplate code for an MCP server.
Curation: Provides a user experience to design the MCP server. This is crucial because a direct one-to-one mapping of an API to MCP tools is often suboptimal. Curation involves selecting which endpoints to expose, remixing multiple APIs into a single server, and refining tool descriptions to improve LLM performance. For example, three separate API calls can be wrapped into a single, more reliable tool like get_all_specs_for_a_company.
Hosting and Management: Offers managed hosting for MCP servers, providing telemetry, performance evaluation, versioning, and security.

This allows companies to quickly launch a modern, agent-ready interface for their services, signaling to developers that their platform supports a “gentic workflows.”

Security, Documentation, and Future Directions

Security: MCP security fundamentally relies on strong underlying API security. OAuth is becoming the standard for authentication. As prompts can be a new vector for attacks, there is a growing need for tooling to perform security and penetration testing on MCP servers.
Documentation and Discovery: MCP is not just for actions but also for discovery. Companies are creating MCP servers for their documentation, allowing developers to ask questions about an API (“how do I use this SDK?”) directly from their IDE without context switching.
Future Tooling Needs: There are still gaps in the ecosystem. Foundation models need to improve their reliability in tool calling, especially with a large number of tools, and better handle API features like pagination.
Edge and Local Use Cases: The MCP standard is flexible. An agent on an edge device (like a phone) could interact with a local MCP server (e.g., pointed to localhost) to access local data like photos, even when offline.