What Does Your MCP Server's Tool List Cost an AI Agent?
What does your MCP server's tool list cost an AI agent? The typical menu, the heavy ones, and why nobody on your team notices the cost.
Every time an AI agent connects to your MCP server, it first loads your full tool list, with every name, description and input schema, and it pays for that in tokens before it takes a single useful step. Across the public servers 7Maps checks, the typical tool list costs about 1,700 tokens. Some cost over 25,000 tokens, about 15 times the typical one. The server keeps working the whole time, so nobody on the team sees the cost. The people who see it are the ones running the agent, and they decide which servers are worth connecting to.
The short version
- The tool list is the menu. An agent reads all of it before it orders anything.
- The menu is paid for on every connection. It is a cost of using your server, not a one-time setup.
- Weights vary widely. The typical server is light; a few are many times heavier.
- Nothing looks broken. A heavy menu passes every uptime check and every test.
- The cost lands on whoever runs the agent. That person chooses whether your server stays connected.
What is an MCP tool list, and why does an agent load all of it?
The Model Context Protocol lets a company expose its product to AI agents as a set of tools. When an agent connects, it asks the server which tools it has, and the server answers with the full list, as the MCP specification for tools describes. The agent needs that list to decide what to call, so it places the whole thing in its working context. Every word in a description, every parameter in a schema, is text the agent reads and pays for.
How heavy is a typical MCP tool list?
The live numbers are on the 7Maps state page, updated daily, and they count only servers that answered with a tool list. The gap matters at volume: an agent that connects 1,000 times a day to a server 23,300 tokens heavier than the typical one reads about 23 million extra tokens a day before doing any work, and pays for each of them at its model's price.
Our own server. Our own full menu is about 8,500 tokens because it carries 16 tools with payment and safety options; the 9-tool agent version of 7Maps is about 2,100, near the median. We measure ours the same way.
Why doesn't anyone on your team notice?
Because nothing fails. The server answers, the tools work, the demo looks good and the monitoring stays green. The weight shows up somewhere else: in the token bill of the company running the agent, and in an engineer on that side deciding which connectors to keep. Teams that build an MCP server usually test that it works. Very few test what it costs to use, and standard observability for small apps watches errors and uptime, not menu weight.
What does a heavy tool list mean for the business?
- The bill lands on whoever runs the agent. If that is a customer, your product just became more expensive to use.
- Every request carries the weight again. The full menu is read on every connection, before the first useful step.
- It competes for room. An agent connected to several servers has a limited working context, and a heavy menu takes a larger share of it.
- It is invisible in your own reports. Uptime, error rates and usage all look normal.
How can you tell if your tool list is heavy?
- Count the tools your server lists on a fresh connection.
- Read the longest description. If it reads like a full user manual, the menu is carrying documentation.
- Look for near-twins, several tools that do nearly the same thing.
- Compare releases. A tool list that grows with every release and never shrinks is getting heavier.
These signs tell you whether to look closer. How to fix it depends on the server, and that is work worth doing with care, because a careless cut can remove what an agent needs.
Common pitfalls
The most common one is treating the tool list as documentation. Teams write it for a human reader and keep adding detail, which makes sense for docs and costs money in a menu that is read on every connection. The second is measuring only once at launch. Tool lists change with releases, and a server that was light in spring can be heavy by autumn without anyone deciding it. It builds up the way technical debt does, one reasonable change at a time.
A typical path: a company launches an MCP server with a handful of tools, adds features every few weeks, and copies its help pages into the descriptions. A few months later the menu is many times its launch weight, and every test still passes.
When is it worth getting a senior review?
When your MCP server touches revenue, when customers run agents against it at volume, or when the tool list has grown release after release and nobody owns its size. We are an AI software company that builds AI systems for businesses, and reviewing and improving MCP servers is part of what we do. Related reading: structured output and tool calling, running an LLM in production and, if you also run a store or booking site, five checks for whether AI assistants can find it.
Free check
Look up your server on the map and see how it compares: 7it.co.il/7maps.
FAQ
What is a token in this context?
A token is the unit AI models use to measure and price text. More words in your tool list means more tokens read on every connection.
Does a heavy tool list break anything?
No, and that is the problem. It raises cost and slows responses for the people running the agent, while every check on your side stays green.
Is the number of tools the main factor?
Not always. A server with few tools can still be heavy when each description is very long, and a server with many short tools can be light.
How often does a tool list change?
It changes whenever the server ships new or edited tools, which for active products can be every few weeks. Its weight is worth checking after each release.