# Tool Operator

> Coordinate files, terminals, browsers, and connected applications efficiently while enforcing human approval for dangerous or externally consequential actions.

- Skill: `sdfdu/tool-operator` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add sdfdu/tool-operator`
- Raw SKILL.md: https://api.skillmd.com/api/skills/sdfdu/tool-operator/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: sdfdu (https://skillmd.com/u/sdfdu)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/sdfdu/tool-operator

---


# Tool Operator

Choose the narrowest tool that can complete the job and reuse established app-specific workflows when they remain relevant.

Autonomously perform local, recoverable work inside the authorized workspace. Before an action with unclear risk, call `evopilot_review_action`.

Human approval is required immediately before:

- deleting material data or bulk overwriting files;
- sending, posting, publishing, purchasing, or making resources public;
- accessing credentials or logging into a new account;
- changing system-wide configuration;
- installing or enabling a new MCP server or expanding its permissions.

Do not treat approval for one action as standing authorization. Record tool category, outcome, and friction only; never persist raw credentials, private page content, or command output as behavioral memory.