MCP Tool Description Linter
Lint your MCP tool definitions against Anthropic's agent-tool best practices: naming, description quality, parameter docs, and context-window cost per loadout — all in your browser.
MCP Tool Definitions (JSON)
Lint Report
About this linter:
Rules distill Anthropic's "Writing effective tools for agents" guidance, the MCP specification and the rubric from arXiv:2602.14878. Token counts use the cl100k encoding locally in your browser — nothing you paste is ever sent anywhere.
How to Use the Tool
Tool descriptions are the only thing an LLM agent sees when deciding which tool to call. This linter checks your MCP tool definitions against the practices that measurably improve tool-selection accuracy.
- Export your tools: From your MCP server, serialize the tool list — a single tool, an array, or the tools/list response envelope.
- Paste the JSON: Linting runs instantly: naming, description quality, schema completeness and per-tool token cost.
- Fix the findings: Errors break tool discovery; warnings degrade selection accuracy; notes are polish suggestions.
- Watch the loadout cost: The per-tool breakdown shows how many tokens each description and schema consumes of the context window every turn.
Frequently Asked Questions (FAQ)
Why do tool descriptions matter so much for agents?
The model never sees your implementation — only the tool name, description and input schema. Anthropic's engineering guidance shows that rewriting descriptions to state what a tool does and when to use it is one of the highest-leverage fixes for agents picking the wrong tool.
Which input formats can I paste?
Three shapes are recognized automatically: a single tool object with a name, a description and an inputSchema, an array of tool objects, and the tools/list envelope with a tools array inside.
How are tokens counted?
Locally in your browser with the cl100k encoding (the family used by GPT-4-class models). Claude and Gemini use different tokenizers, so treat the numbers as a comparable relative cost, not an exact invoice.
What is a good score?
100 means every rule passed. In practice, aim for zero errors and zero warnings — the remaining notes (like adding "use this when" guidance) are incremental improvements.