# MCP Tool Definition Token Overhead Calculator

> Source: https://netray.co/tools/mcp-tool-definition-context-token-overhead-calculator
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Sizes the context window and input cost eaten by MCP tool schemas before the user says a word, and shows the what-if as you add tools.

## Category

developer

## Audience

- AI engineer
- ERP integration developer
- Solution architect

## Tool

**Kind:** calculator

**Currency:** $

**Inputs:**

#### Number of tools exposed

**Key:** tools

**Type:** slider

**Default:** 30

**Min:** 1

**Max:** 150

**Step:** 1

**Unit:** tools

**Help:** Total tools across all connected MCP servers that the client sends to the model.

#### Average schema size per tool

**Key:** chars_per_tool

**Type:** number

**Default:** 900

**Min:** 100

**Step:** 50

**Unit:** characters

**Help:** Name, description and JSON input schema serialized. ERP tools with many filter fields run 600 to 1,500.

#### Characters per token

**Key:** chars_per_token

**Type:** number

**Default:** 3.5

**Min:** 2

**Max:** 6

**Step:** 0.1

**Help:** JSON and code tokenize worse than prose. 3 to 4 is typical for English JSON.

#### System prompt and other fixed tokens

**Key:** system_tokens

**Type:** number

**Default:** 1500

**Min:** 0

**Step:** 100

**Unit:** tokens

#### Model context window

**Key:** window

**Type:** number

**Default:** 128000

**Min:** 4000

**Step:** 1000

**Unit:** tokens

#### Model calls per day

**Key:** requests_day

**Type:** number

**Default:** 4000

**Min:** 0

**Step:** 100

**Help:** Every agent step re-sends the tool list unless it is cached.

#### Input price per million tokens

**Key:** price_in

**Type:** number

**Default:** 3

**Min:** 0

**Step:** 0.1

**Unit:** $ per 1M

**Outputs:**

#### Tool schema tokens

**Key:** schema_tokens

**Expr:** tools * chars_per_tool / chars_per_token

**Format:** integer

**Primary:** true

**Description:** Tokens spent on tool definitions on every call.

#### Total fixed prefix

**Key:** fixed_tokens

**Expr:** schema_tokens + system_tokens

**Format:** integer

#### Share of context window

**Key:** pct_window

**Expr:** fixed_tokens / window * 100

**Format:** percent

**Primary:** true

#### Tokens left for task and history

**Key:** left

**Expr:** max(0, window - fixed_tokens)

**Format:** integer

#### Daily input cost of the prefix (uncached)

**Key:** daily_cost

**Expr:** fixed_tokens * requests_day / 1000000 * price_in

**Format:** currency

#### Monthly input cost of the prefix (uncached)

**Key:** monthly_cost

**Expr:** daily_cost * 30

**Format:** currency

**Chart:**

#### Fixed prefix share of the window as tools are added

**Type:** line

**X:** tools

**From:** 1

**To:** 150

**Steps:** 30

**Series:**

- pct_window

**Disclaimer:** Token counts are estimates from a characters per token ratio. Use your model's tokenizer for exact counts and your provider's price sheet for cost.

## Content

**Intro:** Every MCP tool you expose is serialized into the model's context on each call. With a few dozen ERP tools the schemas alone can cost thousands of tokens before any user text. This calculator turns tool count and schema size into tokens, window share and money.

**Sections:**

#### How to use it

**Body:** Enter the total number of tools the client will send, not just one server's count. Paste a representative tool definition into a character counter to get the average size, then set characters per token for your model. The primary outputs are the schema tokens and the share of the context window they occupy. Drag the tools slider to see where the prefix crosses a share you are not willing to pay on every call. The chart sweeps tool count from 1 to 150 so you can find the knee in the curve before you design a server with sixty fine-grained tools.

#### The math behind it

**Body:** Schema tokens equal tools times average characters divided by characters per token. The fixed prefix adds the system prompt. Window share is prefix over window. Cost is prefix tokens times calls per day divided by one million times the input price, times thirty for a month. The estimate ignores output tokens and any prompt caching. If your provider supports caching of a stable prefix, the cost line drops sharply for cache hits, which makes a stable, sorted tool list worth keeping.

**Bullets:**

- Long descriptions and enum lists dominate the character count.
- Tool order changes can break prefix caching.
- Selection quality tends to degrade as the list grows, not only cost.

#### Reducing the overhead

**Body:** Merge near-duplicate tools into one with an operation argument, shorten descriptions to what the model needs to choose correctly, and remove optional parameters nobody uses. For ERP work, expose a small set of query and action tools rather than one tool per table. If the client supports it, load tool groups on demand per task. Re-run this calculator after each change and compare the monthly line, not just the percentage.

**Cta:** If your ERP MCP server has grown past what one context can carry, Netray can help restructure the tool surface and run it against private models on your own hardware.

**Faq:**

#### How many MCP tools is too many?

**Answer:** There is no fixed limit. Practically, watch two numbers: the share of the context window taken by schemas and the model's tool selection accuracy in your own tests. Many teams find quality slips noticeably once the list reaches several dozen overlapping tools.

#### Does a tool that is never called still cost tokens?

**Answer:** Yes. Unless the client loads tools lazily, every connected tool's definition is sent on each model call, used or not. That is why trimming unused tools lowers cost directly.

#### Is characters per token 4 accurate?

**Answer:** It is a rough prose figure. JSON with braces, quotes and identifiers usually tokenizes at roughly 3 to 4 characters per token, so 3.5 is a safer default. Run a few real schemas through your tokenizer to calibrate.

#### Does prompt caching fix this?

**Answer:** It reduces the price of re-sending a stable prefix, but not the space it takes in the window. A large tool list still limits room for history and results, so size matters even with caching.

## Search terms

mcp tool schema token count, mcp context window overhead, too many mcp tools, tool definition tokens llm, mcp server token cost, how many tools can an llm handle, function calling schema token calculator

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json