Thoughts on the future of MCP design

In the age of AI, a lot of MCP design is already heading in the wrong direction — giving models more tools is not the same as making them more capable.

More tools is not more capability

Giving an LLM access to 50 tools doesn't automatically make it more capable. If every tool dumps raw data back into the context and expects the model to calculate, filter, sort, validate, and figure out what matters, you're just moving backend work onto the LLM.

Do the reliable work before the model sees it

If something can be calculated or determined reliably with code, it probably should be. The MCP layer should do that work before the result ever reaches the model — trends, deduplication, ranking, and known API shapes included.

Boring tools, better judgment

The best MCP tools will feel boring: narrow responsibilities, predictable inputs, strong validation, and exactly what the model needs for the next decision. Code handles precision; the LLM handles ambiguity, reasoning, and judgment.

All posts · Home