Half the posts about "discovering MCP servers" stop at the list: registries, directories, search endpoints. Useful, but discovery is only the first half of the problem. Every server you connect after discovering it pays rent in your context window, and nobody shows you the rent before you sign.
One catalog, two prices. Native tools/list schemas across a 50-server,
255-tool corpus versus mcptoon's names-only index. tiktoken cl100k_base, source of truth
assets/benchmark_tiktoken.json. −99.2%, every tool still callable.
Find candidates. Registries do this well. Search a directory, read the descriptions, star the promising ones. Ten minutes, done. This part is solved.
Decide what they cost. This half has no tooling, which is why people connect everything and wonder why their agent got slow, and it's also why "add more servers" advice keeps floating around without a word about what the additions do to your window. A server with 40 tools costs 40 schemas, every turn, in every client that reads full catalogs. The discovery question isn't "does this server exist." It's "does this server earn its tokens." Most tooling never asks the second question. We built the part that does, and honestly the second question is the only one with money attached to it.
Not a web registry. Your machine. It scans local configs and running processes for MCP servers you already have, connects, and reports what each catalog costs in tokens. Then it connects them all behind one name index so the cost stops mattering. The flow:
python -m pip install mcptoon mcptoon discover # find servers already on this machine mcptoon quickstart # pick them, write one config mcptoon sync # push it to every agent you use
After that, your clients see one short index instead of fifty long catalogs, and the price of adding an eleventh or a twentieth server stops being a number you have to think about at all, which is the quiet part that makes the whole setup sustainable. The token tax calculator prices any count in the browser if you want the number before installing. The client-side story is in the Cursor guide; the protocol mechanics are in the tools/list token bloat guide.
| Tool type | Finds servers? | Prices them? |
|---|---|---|
| Web registries / directories | yes | no |
| Client built-in marketplace | yes | no |
| mcptoon local discover | yes, already installed | yes, per catalog |
Skill catalogs have the same missing price tag, and once you see it you can't unsee it: 489 skills cost 1,572,330 tokens of catalog until a resident pointer cut them to 39. Same disease, same fix, different asset class. Whatever you're discovering, ask the rent question first.
Is there an official MCP server registry?
The ecosystem has several directories and client built-in marketplaces, and they change often. We won't pretend one canonical list exists. What doesn't change is the cost side: whatever registry you use, the servers you connect bill your context by schema.
How many servers should I connect?
As many as you actually use. The failure mode isn't the count. It's paying full schema price for all of them. With a name index in front, 50 servers and 255 tools cost 581 tokens. That changes the answer from "three, carefully" to "everything you're really using." Here's the check: if you can't say what a server is for, it shouldn't be connected regardless of the price.
Does discovery send my data anywhere?
No. Discovery reads local config files and running processes on your machine. Nothing uploads. Nothing leaves your machine. It's the same reason the calculator runs in your browser.