What jumped out at me isn’t that MCP has overhead. Of course it does. It’s that this post turns “MCP is chatty” into a very crisp bill of goods, and some of that crispness makes me squint.
The basic complaint is believable: if you hand an agent a pile of tool schemas every time, you are paying to restate a lot of structure before you get to any useful work. That part is not controversial. JSON schemas are verbose, tool wrappers are verbose, and repeated discovery costs tokens. If you’ve ever watched an agent spend half its context on scaffolding, you already know the shape of the problem.
What I’m less convinced by is the exactness of the numbers. “91,247 tokens,” “812 per call,” “60% of your tokens are JSON braces” — those are the kind of figures that feel authoritative right up until you ask how they were measured, with what tokenizer, in what conversation shape, and whether every server really behaved the same way. The post says the counts were measured, not estimated, which is helpful, but the article doesn’t really show enough methodology for me to treat the totals as universal. I’d want to know whether Claude Code is actually reinjecting all 255 tools every time, or whether this is a specific setup that happens to be especially bad.
The other thing that makes me hesitate is the comparison to TOON. Maybe it’s a good compression trick. Maybe it really is useful for schema-heavy tool lists. But “62 tokens instead of 2,034 for all 255 tools” is exactly the sort of claim that needs a closer look. Compression formats tend to look fantastic when you test them on the thing they were designed to flatten. The harder question is whether they stay readable, debuggable, and compatible once you leave the happy path. Agents are already fragile enough without adding another translation layer in the middle. I think that’s the trade: token savings on one side, operational complexity on the other.
Still, I don’t think the post is wrong to poke at MCP here. If a tool protocol becomes popular enough that people start wiring up dozens of servers, the costs stop being theoretical. At that point, schema design is not just an API concern; it is a context-budget concern. That’s a useful reminder, and honestly the most interesting part of the piece. Not “MCP is doomed,” which I don’t buy, but “protocols for agents have to care about prompt weight in a way ordinary APIs usually don’t.”
What I would actually try first is simpler than the author’s full proxy story: prune tool sets aggressively, shorten descriptions, and stop exposing giant tool catalogs to every conversation. If that still feels expensive, then yes, I’d test a proxy or schema cache. But I’d want evidence from my own workload before I decided the answer is a new notation format.
Reference: Real Token Cost of MCP: 91K Tokens of JSON
