Every MCP server you connect brings its tool list. Every tool's schema lands in the model's context. Ten servers later, the agent reads dozens of schemas, often hundreds, before answering a single question. That's tool overload, and it has a price per turn.
A 50-server, 255-tool catalog. Native tools/list with full JSON schemas on
the left; mcptoon's names-only index on the right. tiktoken cl100k_base, source of truth
assets/benchmark_tiktoken.json. That's −99.2%, every tool still callable.
Not a crash. Not an error message. The failure is quieter: the model spends its attention on schemas instead of your problem, tool selection gets worse, sessions fill up sooner, latency creeps in a way you can feel but can't quite prove, and the cost renews every single turn whether or not the agent ever touches those tools.
The protocol isn't broken. MCP has no "summary mode" for tools/list, so clients that
want full schemas get full schemas. At three servers that's fine. At thirty it's the biggest standing
cost in your agent's context, and it's bigger than most conversations you'll ever have with the thing.
You'll feel these before you measure them.
search versus search_web versus
web_search versus whatever your search server actually named it, and once the confusion
starts it compounds, because a wrong pick teaches the model the wrong lesson about the catalog.| Approach | Effect | Trade-off |
|---|---|---|
| Disconnect servers by hand | small | you lose tools, churn follows |
| Hope the model "copes" | none | the cost is paid before selection |
| A gateway serving one name index | −99.2% | one lookup step, handled for you |
mcptoon fronts your servers. Schemas stay on disk. The agent sees one compact name index and pulls the one schema it needs, on demand. Four commands to see it on your own machine:
python -m pip install mcptoon mcptoon demo # token math, live, zero config mcptoon quickstart # discover servers, write one config mcptoon sync # push it to every agent you use
The same gateway covers skill catalogs, which fail the same way for a different reason: 489 skills cost 1,572,330 tokens until a resident pointer cut them to 39. And if you're on Cursor specifically, the client-side view is in the Cursor guide. Prefer to price before installing? The token tax calculator runs in your browser.
How many MCP tools is too many?
There's no protocol limit. The practical ceiling is your context window, and the cost curve is linear in schemas: 255 full schemas cost 71,929 tokens here. Past a few dozen tools, overhead starts crowding out work.
Doesn't the model just ignore irrelevant tools?
It tries. But ignoring isn't free: the schemas are already in context, already paid for, and they still compete for attention during selection. Overload degrades picks before the model can opt out.
Is this an MCP protocol flaw?
No. Full schemas up front is a reasonable default for small setups. It just scales badly, which is why the fix lives in a gateway rather than a fork of the protocol.