Most MCP servers are a thin wrapper around readFile and grep. That is easy to write and expensive to use: ask an agent to add a field to a service and it will grep up 240 matching lines that are mostly imports, read 4,100 tokens of a file to use about 30 of them, and then read 3,900 lines of test log to find four failures. Roughly 14,000 tokens to learn four facts, and the context window is now full of material it will never look at again.
Jellybean does the same task in under 900 tokens, and the agent ends up knowing more — because it also knows what else is in those files and what depends on them.
It comes down to two decisions. Results are handles, not dumps: every row carries an identifier addressing an exact region of an exact file, and the agent spends tokens on a body only for the handles it chooses to follow. Handles are derived from content, so the same region always produces the same one, and an agent can tell a search hit and an outline entry are the same function without expanding either.
And the token budget is a contract rather than a hint. Every tool takes one and every tool honours it.