MCP Tool Schema Token Overhead Calculator

Every tool or MCP server you expose adds a definition (name, description and parameter schema) that the model has to read, usually on every request. Enter how many tools you have, the average tokens per definition and any fixed protocol overhead to see how much of the context window they use, what is left and what they cost across sessions. You can also paste your real definitions and measure them in the browser; nothing you paste is uploaded.

Processed locally in your browser. Your data is not uploaded to our servers.

Tools or functions exposed to the model.

Name, description and parameters together. Many definitions run from about 100 to a few hundred tokens: measure yours below.

Instructions or protocol text added around the definitions. 0 if none.

Measure pasted tool definitions (optional)

Paste the JSON of your tool definitions. They are measured in your browser, never uploaded and never sent to analytics.

The text is only counted: it is not parsed as code or run.

Each request normally carries the definitions again.

How many sessions use these tools in a day.

Check your model’s documentation for its context window. Input and output usually share it.

Leave empty to see tokens only.

30 by default.

Result

Estimate

Enter the tools, tokens, turns and context window to see the overhead.

Exact arithmetic on your counts. Real clients and models may serialize and count tools differently, so treat the figures as an estimate and compare with a real request’s usage.

Examples

Worked examples (128,000-token window, 10 requests per session, 200 sessions a day, example price of 3 USD per million input tokens):

SetupOverhead per requestShare of the windowTokens per dayCost per month
Small: 5 tools9000.7%1,800,000162.00 USD
Typical: 20 tools3,5002.7%7,000,000630.00 USD
Large: 100 tools20,80016.3%41,600,0003,744.00 USD

Formula & Methodology

The arithmetic is exact for the numbers you enter:

overhead per request = tools × schema tokens per tool + fixed overhead (or the measured total + fixed overhead)share of the window = overhead per request ÷ context window; remaining = window − overhead per requestoverhead per session = overhead per request × requests per sessioncost per request = overhead per request × input price ÷ 1,000,000; per day = cost per request × requests per session × sessions per day

Good to know

  • Clients differ: some send every tool definition on every request, some load tools on demand or cache them. This tool assumes the definitions are sent with each request, so adjust the requests per session to match your client.
  • Vendor-specific accounting is not universal. A provider may add its own tool-use system tokens, and tokenizers differ, so a measured count is only exact for the encoding named.
  • Measured JSON text is estimated at about 4.5 characters per token (range 2.5 to 5, from OpenAI tokenizers measured on tool definitions) unless you pick an exact OpenAI encoding. Real tokenizers vary.
  • Prompt caching can make repeated definitions cheaper. This tool bills the overhead at the plain input price: see the prompt cache calculator.
  • Fewer, better-described tools usually cost less context and help the model choose. Measure before and after.

Method and references

These primary sources support the methods this tool relies on. Having read them does not verify every result the calculator produces, and figures listed as assumptions are ToolQuarter’s own.

The structure of a tool definition comes from the MCP specification. How many tokens it costs depends on the client, the model and the tokenizer.

  • MCP specification: Tools (2026-07-28): Defines what a tool definition contains: a name, an optional title, a description, an inputSchema written in JSON Schema and an optional outputSchema. It describes structure, not a token count. ()
  • OpenAI API: Function calling: States that function definitions count against the model’s context limit and are billed as input tokens, and suggests loading fewer functions and shortening descriptions. ()
  • Anthropic: Pricing, tool use: Shows that another provider also bills the tools parameter as input tokens and adds its own tool-use system prompt tokens per model, which is why overhead differs between providers. ()

ToolQuarter assumptions (no primary source):

  • Measured JSON is estimated at about 4.5 characters per token (range 2.5 to 5) unless you choose an exact OpenAI encoding. That range is ToolQuarter’s own measurement on tool definitions, not a provider figure.
  • The calculator assumes the client sends every definition with each request. The sources read do not say how often a client does this and clients differ (some load tools on demand), so adjust the requests per session.

How to Use

  1. Enter the number of tools, the average schema tokens per tool and any fixed overhead.
  2. Optionally paste your tool definitions to measure them locally and use the measured total.
  3. Enter the requests per session, the sessions per day, your context window and, if you like, your input price.
  4. Read the overhead, the share of the context window, what remains and the cost.

Frequently Asked Questions

How many tokens do MCP tool definitions use?

It depends on the tools. A definition with a name, a description and a parameter schema is often around 100 to a few hundred tokens, so twenty tools can take a few thousand tokens. Paste your own definitions to measure them rather than relying on an average.

Are tool definitions sent on every request?

In many clients, yes: the model has to see the definitions each time. Some clients load tools on demand or cache them. This calculator assumes they are sent with each request, which you can adjust through the number of requests per session.

Is the text I paste uploaded anywhere?

No. It is measured in your browser and never uploaded, parsed as code or sent to analytics. If you choose an exact encoding only the tokenizer’s data is downloaded.

Is the count exact for my model?

Only for the OpenAI encoding you choose, and only for the text as you pasted it. Other models tokenize differently and clients may serialize tools in their own format, so treat it as an estimate.

How can I reduce the overhead?

Expose fewer tools, shorten descriptions and schemas, load tools only when needed and use prompt caching for definitions that repeat. Then measure again.

Last updated: