Why seven tools is the right number

You've wired in 30 MCP tools. Maybe 50. Each one gets a name, a description, and a JSON schema. The model reads all of them on every single call.

That's the cost nobody talks about.

The token tax you're already paying

Every MCP tool definition costs 50–150 tokens in the model's system prompt. The name, the description, the input schema — it all goes in.

Ten tools? About 1,000 tokens. Manageable.

Fifty tools? You're spending 5,000+ tokens before the model thinks a single thought. That's input tokens on every call, every turn, every retry. It compounds fast.

But the input tokens aren't even the expensive part.

The reasoning tax is worse

The model has to look at every tool in the list to decide which one to call. More tools means more output tokens spent reasoning through choices. And output tokens cost 3–5x more than input tokens.

A model staring at 50 tools and picking the wrong one wastes a full round trip — the failed call, the error handling, the retry. That's not a rounding error. That's real latency and real money on every mistake.

The research backs this up. Tool selection accuracy drops as the number of available tools increases. It's not a subtle effect. More choices means more wrong choices.

What if the list never grew?

That's the core idea behind mcp-flowgate. Instead of registering every capability as a separate tool, you expose exactly seven:

• Three for discovery: gateway.home, gateway.search, gateway.describe
• Four for action: workflow.start, workflow.get, workflow.submit, workflow.explain

Your 50 capabilities still exist. They just surface through search results and response links — loaded one at a time, only when relevant.

The model's tool list stays at seven entries whether you have 5 capabilities or 500.

Search beats scanning

When the model needs to publish content, it doesn't scan 50 tool definitions looking for the right one. It calls gateway.search with "publish content" and gets back one result.

One search call. One result. The model follows the link to start the workflow. No scanning, no guessing, no wasted reasoning.

This pattern has a name in API design: HATEOAS. The server tells the client what's legal next. The model doesn't need out-of-band knowledge. It just follows the links.

Seven is a ceiling, not a target

We didn't pick seven because it's a magic number. We picked it because it's the minimum set that covers the full lifecycle: discover what's available, act on it, and inspect the rules.

Discovery needs three tools because browsing, searching, and reading details are different operations. Action needs four because workflows have state: you start them, check them, advance them, and sometimes need to understand the rules before deciding.

Seven tools. Fixed token cost. And every capability you add behind them comes with governance — guards, audit, retries, approval gates — built in.

Try it yourself

The quick start takes 30 seconds. Wire in your tools, point the model at mcp-flowgate, and watch the tool list shrink while your capabilities grow.