🗜️

MCP Tool Description Compressor

Shrink MCP tool descriptions and JSON schemas to cut context cost — without changing how agents call them.

TOKEN SAVINGS

Der RechnerPage 1

Tool definitions (JSON array)

[
  {
    "name": "search_customers",
    "description": "This tool allows you to search through the customer database in order to find matching records. It is able to handle a large number of different query formats, including partial matches (e.g. \"john\" matching \"John Smith\"). Please provide at least one search field. Results are paginated and sorted by relevance by default.",
    "inputSchema": {
      "type": "object",
      "required": [
        "query"
      ],
      "properties": {
        "query": {
          "type": "string",
          "description": "The search term to look for across customer name, email and company fields."
        },
        "limit": {
          "type": "number",
          "default": 25,
          "examples": [
            10,
            25,
            50
          ],
          "description": "Maximum number of records to return in a single page of results."
        },
        "status": {
          "type": "string",
          "enum": [
            "active",
            "churned",
            "trial"
          ]
        }
      }
    }
  },
  {
    "name": "refund_order",
    "description": "This function issues a refund for a given order. Use this tool to reverse a charge after support has approved it. Please note that partial refunds are supported by passing an amount lower than the original total.",
    "inputSchema": {
      "type": "object",
      "required": [
        "orderId"
      ],
      "properties": {
        "orderId": {
          "type": "string"
        },
        "amount": {
          "type": "number",
          "default": null,
          "description": "Optional partial refund amount in the smallest currency unit. Omit to refund the full order total."
        },
        "reason": {
          "type": "string",
          "enum": [
            "duplicate",
            "fraudulent",
            "requested_by_customer"
          ]
        }
      }
    }
  }
]

Each item: {name, description, inputSchema}

Tokens before

330

Tokens after

130

Reduction

60.6%

Tool	Tokens before	Tokens after	Saved	Reduction	Description chars
search_customers	186	66	120	64.5%	322 → 76
refund_order	144	64	80	55.6%	212 → 34

TOKEN MODEL: description at 4.0 chars/token · schema at 3.6 chars/token

Compressed definitions — ready to paste

[
  {
    "name": "search_customers",
    "description": "allows you to search through the customer database to find matching records.",
    "inputSchema": {
      "type": "object",
      "required": [
        "query"
      ],
      "properties": {
        "query": {
          "type": "string"
        },
        "limit": {
          "type": "number"
        },
        "status": {
          "type": "string",
          "enum": [
            "active",
            "churned",
            "trial"
          ]
        }
      }
    }
  },
  {
    "name": "refund_order",
    "description": "issues a refund for a given order.",
    "inputSchema": {
      "type": "object",
      "required": [
        "orderId"
      ],
      "properties": {
        "orderId": {
          "type": "string"
        },
        "amount": {
          "type": "number"
        },
        "reason": {
          "type": "string",
          "enum": [
            "duplicate",
            "fraudulent",
            "requested_by_customer"
          ]
        }
      }
    }
  }
]

Descriptions are reduced to their first sentence; property-level defaults, examples and descriptions over 60 characters are dropped. Enum values and required lists are always preserved — call semantics do not change. Atlassian’s mcp-compressor reports 70–97% description-overhead reduction without changing how agents call tools; a typical safe reduction is 40–70%. Re-test your agent after swapping in the compressed definitions.

Data Source & Legal Disclaimer

Effective: Token model: description at 4.0 chars/token, schema at 3.6 chars/token

Sources: Compression model mirrors the FreeToolHub MCP tool compress_tool_definitions

Token counts are estimates from a character-based model, not measured tokenizer output. Always re-test your agent after compressing: this tool preserves enum values and required lists, but a description you trimmed to one sentence may have carried nuance your model relied on to choose correctly.

See all data sources & update policy →

So funktioniert esPage 2

↩ Zurück zum Rechner

Most of your context budget is being spent on prose nobody reads

Tool descriptions are written for humans browsing a README, then loaded verbatim into every single request. The waste is remarkably consistent: hedging(“This tool allows you to…”), restatement (a sentence explaining what the function name already says), examples in prose, and schema decoration — default, examples and paragraph-long property descriptions that the model rarely needs to select a tool. None of it changes how the tool is called, and all of it is re-sent on every turn. The key safety property here: enum values and required lists are never touched. Those define the call contract, so stripping them would break invocation rather than merely shorten it. What gets cut is prose and decoration — which is why a typical safe reduction lands in the 40–70% range without a behaviour change.

What survives compression, and what is dropped

Type information, enums and required lists define how the model calls your tool, so they are always preserved. Descriptions, defaults, examples and long property text only help a human reader, so they are reduced or removed.

Two realistic tools, one pass

A server exposes search_customers and refund_order. Both descriptions were written for a README: they hedge, restate the function name, and inline examples. Both schemas carry defaults and paragraph-length property text.

Descriptions:Hedging stripped, examples removed, each cut to its first sentence — the operative instruction survives
Schemas:default and examples dropped; property descriptions over 60 chars dropped
Preserved:type, required, per-property type and every enum — the call contract is byte-identical
Result:Roughly half the tokens, and the agent calls the tools the same way

↩ Zurück zum Rechner

MCP Tool Description Compressor

Most of your context budget is being spent on prose nobody reads

Keep the Free Edition Free