Anthropic admits that MCP sucks

The speaker introduces MCP as an example of AI being a bubble, noting a disproportionate number of companies building observability tools for MCP compared to those making useful products with it. He criticizes MCP for requiring excessive context, making models dumber and less efficient. He points out that Anthropic's new article, "Code Execution with MCP," implicitly acknowledges these issues, suggesting that agents should write code to interact with MCP servers rather than using direct tool calls.

The video features a sponsor segment for Savala, a hosting platform. Savala is praised for its ease of use in deploying code, linking to GitHub, creating deployment pipelines, and spinning up databases. The speaker highlights its cost-effectiveness and unique features like one-click CDN enabling through Cloudflare, emphasizing its ability to simplify deployments.

The speaker returns to the Anthropic article, emphasizing their admission that directly using MCP with numerous tool definitions and intermediate results clogs the system prompt, making models less effective and more costly. He explains how MCP forces models to carry the entire context of previous tool calls, leading to massive token consumption and reduced efficiency. He criticizes MCP's lack of features like proper authentication (OAuth) and progressive tool discovery.

Anthropic's proposed solution is to treat MCP servers as code APIs and have agents write code to interact with them. This approach allows agents to load only necessary tools, process data in an execution environment before passing results to the model, and significantly reduces token usage. The speaker highlights a staggering 98.7% reduction in token usage achieved by this method, effectively proving his long-standing criticism of MCP's original design.

The section details the benefits of code execution: improved context efficiency by processing data outside the model, enhanced privacy by preventing sensitive data from entering the model's context, and better state persistence by managing state within the code execution environment. The speaker illustrates these advantages with examples, such as filtering large datasets and handling PII securely, and even mentions how agents can save reusable code as skills, mirroring traditional software development practices.

The speaker observes the ironic cycle of how developers create a standard (MCP), realize its flaws, then essentially reinvent traditional software development practices (like SDKs, documentation, and reusable functions) to make it work. He expresses frustration with this recursive pattern, noting that it reflects a failure by 'LLM people' to grasp fundamental software engineering principles and design functional APIs from the outset. He also briefly discusses the security implications and the need for sandboxed environments for code execution.

The video concludes by reiterating the core argument: software engineers, not AI specialists, should be in charge of defining developer tools and APIs. The speaker asserts that the continuous reinvention of basic software engineering concepts within the AI world proves that developers are indispensable and that AI is not truly replacing them. He emphasizes that the issues with MCP are not novel but have well-established solutions in software engineering, making the protocol's initial design severely flawed.