Azayzel

ToolTax — AI skill for Claude Code

AI community

Find wasted tokens, slow tools, duplicate calls, and expensive MCP servers in AI agent workflows.

How to install ToolTax

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Azayzel/ToolTax and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What ToolTax does

Find wasted tokens, slow tools, duplicate calls, and expensive MCP servers in AI agent workflows.

Alternatives in AI

  • MCP Use — The fullstack MCP framework to develop MCP Apps for ChatGPT / Claude & MCP Servers for AI Agents 9.5k ★
  • Browserwing — BrowserWing turns your browser actions into MCP commands Or Claude Skill, allowing AI agents to control browse 1.2k ★
  • Rulesync — by dyoshikawa - A Node.js CLI tool that automatically generates configs (rules, ignore files, MCP servers, com 909 ★

README

ToolTax

**Find out which AI tools are earning their keep.**

ToolTax audits agent/MCP usage for duplicate calls, failures, empty results, oversized outputs, latency, and low-utility servers. It can analyze existing Claude Code/Codex traces or capture stdio MCP traffic live.

TOOLTAX — AGENT TOOL COST REPORT
====================================================================
Calls: 10   Observed payload: ~35,xxx tok   Estimated waste: ...

SERVER                CALLS     TOKENS    WASTE   SCORE  RECOMMENDATION
----------------------------------------------------------------------------------
context7                   2     ...       ...%    ...   cap or filter tool output
memory                     3     ...       ...%    ...   deduplicate/reuse results
github                     4     ...       ...%    ...   keep
filesystem                 1     ...       ...%    ...   monitor

Why

Adding more MCP servers can make an agent *worse*: more calls, more context, repeated lookups, larger outputs, more latency, and more failure surface. Those costs are usually invisible.

ToolTax makes the visible portion measurable.

What it measures

From saved traces:

  • calls by MCP server/tool
  • argument + result payload size
  • duplicate calls
  • failed calls
  • empty results
  • oversized results
  • end-to-end call latency when timestamps permit
  • a transparent heuristic utility score

From live stdio capture, ToolTax additionally observes:

  • exact proxy-observed call latency
  • exact JSON-RPC request/response byte size
  • advertised tools/list definitions
  • de-duplicated tool schema size across paginated/refresh responses
  • protocol-level and isError tool failures

ToolTax does **not** claim to reconstruct provider billing. Advertised schema tokens and payload tokens use deterministic, provider-neutral estimates so runs can be compared consistently. Live capture preserves the original payload estimates even when the stored trace is redacted.

Quick start

Py