The Cursor tool-limit trap, and the math nobody does

Here's a failure mode I see constantly, and it has a number attached.

You add MCP servers to Cursor one at a time. GitHub, a database, a browser, a docs server, Slack. Each one is useful. Then your agent starts picking the wrong tool, or ignoring tools that are right there, and you assume the model got dumber. It didn't. You blew the tool budget.

The arithmetic

Cursor works best when you keep active MCP tools under roughly 40. Past that, tool selection degrades — the model has too many similar options and chooses badly.

Now the part nobody computes: in this directory, the average server exposes 9.2 tools, the median is 10, and browser/automation servers routinely ship 12–19. So:

• 1 server: ~9 tools — fine
• 3 servers: ~28 — fine
• 5 servers: ~46 — past the wall

You don't hit the limit at fifty servers. You hit it at five. That's why the wall feels like it comes out of nowhere — five useful servers is a completely reasonable setup, and it's already too many tools.

Why "one agent that does everything" is the wrong goal

The instinct is to build a single super-agent with every server attached. The math says don't. Two better patterns:

1. Per-task projects. Cursor reads mcp.json per project. Give your data-analysis project the database servers and nothing else; give your web-research project the search and scrape servers. Each stays well under budget and picks tools cleanly.
2. Disable, don't uninstall. You don't need every tool of every server. Most clients let you toggle individual tools — turn off the ones you never call. A 15-tool server you use three tools of should cost you three, not fifteen.

The principle underneath both: fewer, sharper tools beat a big undifferentiated pile. It's the same reason a focused team beats a committee. The model isn't choosing from a catalog; it's choosing under load, and load is what you're managing.

A 10-second check

Before you add the next server, add up the tools you'll actually have active. If it's near 40, something has to come off first. We built a Config Doctor that does this sum for you — paste your config and it tells you the total and what's over — but you can also just count. The point is to count before the agent starts misbehaving, not after.

The takeaway is almost embarrassingly simple: the average server is 9 tools, the ceiling is ~40, so plan for five servers, not fifty. Do the arithmetic up front and the "Cursor got worse" problem disappears.