MCP has a long way to go, what are the dilemmas?

Words: Haotian

 

After learning, these dilemma analyses of MCP are quite in place, hitting the pain points directly, revealing that MCP has a long way to go, and it is not so easy, so I extend it by the way:

 

1) The problem of tool explosion is real: MCP protocol standards, tools that can be linked are proliferated, LLMs are difficult to effectively select and use so many tools, and no AI can be proficient in all professional fields at the same time, which is not a problem that can be solved by parametric quantities.

 

2) Documentation gap: There is still a huge gap between technical documentation and AI understanding. Most API documentation is written for humans, not for AI, and lacks semantic descriptions.

 

3) The weakness of the dual-interface architecture: As the middleware between the LLM and the data source, MCP has to process both upstream requests and transform downstream data, which is inherently inadequate. When data sources explode, it's nearly impossible to unify processing logic.

 

4) The return structure is very different: the data format is confused due to inconsistent standards, which is not a simple engineering problem, but the result of the overall lack of industry collaboration, which takes time.

 

5) Limited context window: No matter how fast the token limit grows, the problem of information overload always exists. MCP spits out a bunch of JSON data that takes up a lot of context space and squeezes inference power.

 

6) Flattening of nested structures: Complex object structures lose hierarchical relationships in text descriptions, making it difficult for AI to reconstruct the correlation between data.

 

7) Difficulty in connecting multiple MCP servers: "The biggest challenge is that it is complex to chain MCPs together." Although MCP is unified as a standard protocol itself, the specific implementation of each server in reality is different, one processes files, one connects to APIs, one operates databases... When AI needs to collaborate across servers to accomplish complex tasks, it's as difficult as trying to force Lego, bricks, and magnets together.

 

8) The advent of A2A is just the beginning: MCP is just the beginning of AI-to-AI communication. A true AI agent network requires a higher-level collaboration protocol and consensus mechanism, and A2A may just be an excellent iteration.

 

Above.

 

These questions reflect the pains of AI's transition from a "tool library" to an "AI ecosystem". The industry is still in the early stages of throwing tools to AI instead of building a true AI collaboration infra.

 

So, it's necessary for MCP disenchantment, but don't overthink its value as a transition technique.

 

Just welcome to the new world。

 

This article is sourced from Foresight News:

https://foresightnews.pro/article/detail/83584

Respectfully submitted by the AIC Team

May5, 2025