Are MCPs Overhyped? A Rant about MCPs — Henry Mao, Smithery

Key Concepts

• Model Context Protocol (MCP): An open standard for connecting Large Language Models (LLMs) to different services.
• AI Agents: Autonomous entities powered by LLMs that can perform tasks by interacting with various services.
• AI-Native Services: Services specifically designed to be used by AI agents through MCPs.
• Tool Calls: Interactions between AI agents and services via MCPs.
• Agent Experience: The overall experience of an AI agent interacting with services and performing tasks.
• Fragmentation: The increasing number of MCP servers, making it difficult to find high-quality ones.
• High Friction Install: The complex installation process for MCPs, making it difficult for users to set them up.
• Agentic Payments: The process of enabling AI agents to make payments on behalf of users.
• Observability: The ability to monitor and improve MCPs after they have been deployed.

Introduction

Henry, the founder and CEO of Smithery and a member of the MCP steering committee, discusses the current state of the Model Context Protocol (MCP) ecosystem. He provides a high-level overview and addresses the challenges and problems within the space, rather than focusing on technical details.

The Origin Story and Claude's Paradox

Henry recounts his initial interest in the ARC AGI challenge, an IQ test for LLMs. While working on this, OpenAI released GPT-3 and GPT-4, which achieved human-level performance on the challenge. This led to the expectation of widespread autonomous agents. However, in 2025, these agents have not materialized as expected. This is referred to as "Claude's paradox," where advanced AI intelligence is "stuck in a box" and not practically useful. To make AI agents useful, the focus needs to shift to context and capability, specifically the inputs and outputs of the models.

The Promise and Problems of MCPs

In November 2024, Anthropic released the Model Context Protocol (MCP) as an open standard to connect LLMs to different services, aiming to standardize the end-to-end problem of AI agent interaction. While MCPs and a growing ecosystem of services targeting AI agents are promising, new problems have emerged.

User Problems with MCPs

• Fragmentation: The increasing number of MCP servers makes it difficult to find high-quality ones. The MCP committee is working on an official registry to address this, but assigning reputation to MCPs remains an open question.
• High Friction Install: MCPs often have complex installation processes (e.g., five-step installation), making them difficult to set up. This can also lead to the installation of insecure MCPs.
• Agentic Payments: There is no clear plan for handling agentic payments, including how agents can pay on behalf of users and how to avoid subscribing to numerous services.

Developer Problems with MCPs

• Hosting Problems: While Streamlit's HTTP transport has made hosting easier, developers still face challenges with stateful sessions and resumability.
• Lack of Developer Tooling: Existing tools, such as the MCP inspector, are basic. Developers need better tools to design effective MCPs, understand how their tools will be called, and create optimal agent experiences.
• Distribution: Developers need a way to get their MCPs discovered by users.
• Observability: Developers need to be able to monitor and improve their MCPs after they have been deployed.
• Monetization: Developers need a way to make money from their MCPs.

Smithery: An AI Gateway

Smithery was founded in December 2024 to address these challenges. It aims to become the AI gateway to grow and orchestrate the new era of AI-native services for AI agents.

Smithery Playground Demo

A demo of the Smithery playground is presented to showcase what can be built when some of these problems are solved. In the demo, an AI agent is prompted to "find the most pressing issue on my GitHub repository called smidy-CLI and create a new ticket on linear." The agent successfully:

1. Identifies the issue.
2. Calls the "search servers" function to find relevant MCPs.
3. Connects to the best server.
4. Connects to Linear.
5. Accesses GitHub using the GitHub MCP to find bugs.
6. Creates a ticket on Linear with details and a link to the original issue.

This demo illustrates an end-to-end task being solved by an AI agent connected to two different MCPs.

The Future of the Internet

The speaker believes that the future of the internet will be dominated by tool calls rather than clicks. In this new world, the agent experience will be more important than the user experience. This agent experience will be built by a community of developers.

Conclusion

The MCP ecosystem is exciting but faces significant challenges for both users and developers. Smithery is working to address these challenges and enable the growth of AI-native services. The future of the internet will be driven by AI agents and tool calls, emphasizing the importance of creating a seamless and effective agent experience.