Unofficial WHOOP recovery, sleep, strain and HRV for AI agents. Prefer MCP tools if connected; otherwise the package CLI. Use when the user wants WHOOP data or actions through an agent.
Installs into .claude/skills of the current project.
Are you the author of Skill?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/davidmosiah-skill-whoop-mcp)
---
name: whoop
description: >
Unofficial WHOOP recovery, sleep, strain and HRV for AI agents. Prefer MCP tools if connected; otherwise the package CLI.
Use when the user wants WHOOP data or actions through an agent.
---
# WHOOP — skill or MCP
Same binary either way. Do not duplicate the API client.
## Choose a surface
**MCP** — tools appear natively after stdio/HTTP config:
```json
{ "mcpServers": { "whoop": { "command": "npx", "args": ["-y", "whoop-mcp-unofficial"] } } }
```
Do not put mutation flags in that snippet.
**Skill / CLI** — no MCP client required. Same tools:
```bash
npx -y whoop-mcp-unofficial call whoop_connection_status --json '{}'
```
If MCP tools named `whoop_*` are already available, use them. Do not also shell out.
## Loop
1. Call `whoop_connection_status` (or `doctor --json` when that exists).
2. Use read tools as asked.
3. Stop on `USER_ACTION_REQUIRED`. Do not invent env flags. Do not enable mutations from this skill.
## Never
- Paste tokens into git, chat logs, or the prompt
- Copy a mutations-enabled assignment into config