Free tool

MCP Token Calculator

Tool definitions are sent on every request, not once. Paste your MCP server URL to see how many tokens that is, which tools are the heaviest, and what it costs.

Works with streamable HTTP and SSE endpoints, for example https://mcp.example.com/mcp
This server needs authentication
Used for this one request and never stored.

Trimmed the schemas. Now show what is left.

Cutting context cost makes your server cheaper to run. It still does not show a customer what it does. Point MCP Showcase at the same URL for a live playground with per-tool documentation.

How it works

link
Paste the URL

We read the tool definitions the server advertises — the same ones an agent receives.

play_circle
Every tool is measured

Name, description and the full JSON input schema, because all three are sent to the model on every turn.

checklist
See the cost and the culprits

Total tokens per request, an estimated cost per 1,000 turns, and the tools ranked by how much context each one eats.

Tool definitions are a cost you pay on every turn

This is the part that surprises people. An MCP server's tool definitions are not loaded once at the start of a conversation — the model has no memory of them between turns, so the complete list of tools, with every description and every JSON input schema, is sent again with each request. Forty verbose tools is not a one-off; it is a fixed tax on every turn of every conversation your agent will ever have.

Paste an MCP server URL above and this calculator reads exactly what an agent receives, measures it, and shows where the weight actually is.

What it measures

  • Total tokens per request — the number added to your context window before any system prompt, conversation history or tool result.
  • Cost per 1,000 turns — the same figure expressed in money, which is usually what makes the problem land.
  • Tokens per tool, ranked — nearly always a small number of tools account for most of the total, and they are rarely the ones people expect.

How to make it smaller

In order of payoff. Remove tools nothing calls. Servers accumulate them; every one is charged on every turn whether or not it is ever useful. Shorten descriptions to one clear sentence covering what the tool does and when to use it — anything longer is usually written for a human reader who will never see it. Trim the input schemas: long per-property descriptions and deeply nested optional objects are typically the bulk of a heavy tool, and they compound because schemas are verbose JSON rather than prose.

There is a second reason to do this that has nothing to do with money. Tool-selection accuracy falls as the list grows. A model choosing between eighty similarly-described tools picks the wrong one far more often than one choosing between eight — so trimming the list usually makes the agent behave better, not just cost less.

About the estimate

Counts are approximated at roughly 3.6 characters per token, which suits the JSON-dense text of tool definitions and typically lands within about 10–15% of a real tokenizer. That is the right tool for comparing tools against each other and finding bloat. It is not the right tool for reconciling an invoice, and the page says so on every result rather than leaving you to assume otherwise.

Related tools and guides

Fewer, clearer tools need better documentation, not less

Cutting descriptions to save context makes your server cheaper and your agent more accurate. It also makes the server harder for a human to understand, which is a problem the moment you want someone to evaluate it. MCP Showcase resolves that split: the model gets your lean schemas, while prospects get a live playground with generated, readable documentation for every tool — so you are not writing one description that has to serve two very different readers.

Frequently asked questions

Because the model has no memory of them between turns. The full list of tools, with descriptions and JSON schemas, is sent with every single request so the model knows what it can call. A hundred verbose tools is a fixed tax on every turn of every conversation.

It is an estimate, typically within about 10–15% of a real tokenizer. It approximates at roughly 3.6 characters per token, which suits the JSON-heavy text of tool definitions. It is built for comparing tools and spotting bloat, not for reconciling a bill.

Three things, in order of payoff: remove tools an agent will never call, shorten descriptions to one clear sentence of what the tool does and when to use it, and trim the input schemas — long per-property descriptions and deeply nested optional objects are usually the bulk of it.

No. It measures only the tool definitions, which is the part of the context you control by changing your server. The system prompt, the conversation so far and the tool results are all on top of this figure.

There is no protocol limit, but there are two practical ones. Cost grows linearly with every turn, and selection accuracy falls as the list grows — a model choosing between eighty similarly-described tools picks the wrong one far more often than one choosing between eight.

More free MCP tools