Tokens

Analyze token usage of MCP (Model Context Protocol) servers. Helps MCP server authors understand and track how much context window their server consumes.

Why?

MCP servers expose tools, resources, and prompts to AI models. Each tool's name, description, and input schema consumes tokens in the model's context window. As servers grow in complexity, this adds up and affects both cost and available context for actual work.

mcp-tokens lets you:

• See exactly how many tokens each tool contributes
• Track token usage over time with baseline comparisons
• Fail CI builds when token usage increases beyond thresholds

Installation

GitHub Action

See mcp-tokens-action for a GitHub Action that wraps this tool.

From source

cargo install --path .

Pre-built binaries

See Releases for pre-built binaries.

Usage

Basic analysis

# Analyze an MCP server (uses ANTHROPIC_API_KEY from environment)
mcp-tokens analyze -- npx @modelcontextprotocol/server-everything

# Analyze a local server
mcp-tokens analyze -- ./my-mcp-server

Output:

MCP Token Analysis: example-servers/everything v1.0.0
Counter: anthropic (claude-sonnet-4-5-20250929)
============================================================

Total tokens: 1934

Tools (11 tools, 1718 tokens):
----------------------------------------
     653 tokens  zip (desc: 43, schema: 610)
     638 tokens  annotatedMessage (desc: 19, schema: 619)
     629 tokens  sampleLLM (desc: 20, schema: 609)
     ...

Token counting providers

Anthropic (recommended): Uses the Anthropic token counting API for accurate counts. Free to use, requires an API key.

export ANTHROPIC_API_KEY=sk-ant-...
mcp-tokens analyze -- ./my-server

Tiktoken (offline): Uses tiktoken for approximate counts. No API key required, but counts are estimates based on GPT tokenization.

mcp-tokens analyze --provider tiktoken -- ./my-server

CI integration

Generate a baseline on your main branch:

mcp-tokens analyze --format json --output tokens.json -- ./my-server

Compare against the baseline in PRs:

mcp-tokens analyze --baseline tokens.json --threshold-percent 5 -- ./my-server

The command exits with code 1 if the threshold is exceeded:

Baseline Comparison
============================================================

Baseline: 1000 tokens
Current:  1934 tokens
Change:   +934 tokens (+93.4%)

Result: FAILED - Token increase of 93.4% exceeds threshold of 5.0%

JSON output

mcp-tokens analyze --format json -- ./my-server

{
  "counter": {
    "provider": "anthropic",
    "model": "claude-sonnet-4-5-20250929"
  },
  "server_info": {
    "name": "my-server",
    "version": "1.0.0"
  },
  "total_tokens": 1934,
  "tools": {
    "total": 1718,
    "count": 11,
    "items": [
      {
        "name": "myTool",
        "tokens": 653,
        "description_tokens": 43,
        "schema_tokens": 610
      }
    ]
  }
}

CLI Reference

mcp-tokens analyze [OPTIONS] -- <COMMAND>...

Arguments:
  <COMMAND>...  Command to start the MCP server

Options:
  -f, --format <FORMAT>          Output format: text or json [default: text]
  -p, --provider <PROVIDER>      Token counting provider: anthropic or tiktoken
  -m, --model <MODEL>            Model to use for token counting
      --anthropic-key <KEY>      Anthropic API key (or set ANTHROPIC_API_KEY)
  -b, --baseline <FILE>          Baseline JSON file to compare against
      --threshold-percent <N>    Max allowed percentage increase [default: 5.0]
      --threshold-absolute <N>   Max allowed absolute token increase
  -t, --timeout <SECONDS>        Timeout for server startup [default: 30]
  -o, --output <FILE>            Save report to file (for use as baseline)
  -h, --help                     Print help

Environment variables

Variable	Description
ANTHROPIC_API_KEY	Anthropic API key for accurate token counting
MCP_TOKENS_PROVIDER	Default provider (anthropic or tiktoken)
MCP_TOKENS_MODEL	Default model for token counting

Tips for reducing token usage

1. Keep descriptions concise: Tool descriptions are fully tokenized. Be clear but brief.

2. Simplify schemas: Complex nested objects with many optional fields add tokens. Consider splitting into multiple focused tools.

3. Use enums wisely: String enums with descriptions add tokens for each variant.

4. Review schema defaults: Default values and examples in JSON Schema add tokens.

License

Licensed under the Apache License, Version 2.0 <http://www.apache.org/licenses/LICENSE-2.0> or the MIT license <http://opensource.org/licenses/MIT>, at your option.