hevy-coach
MCP server that connects Claude to the Hevy fitness app, enabling AI-driven personalized strength coaching by reading lifting history and programming workouts.
README
claude-hevy-coach
Claude as your personal strength coach, wired into the Hevy app.
Say "coach me" → Claude reads your real lifting history, checks how your last programmed workout actually went, asks how you're feeling, designs your next session, and writes it into Hevy as a routine. At the gym you just open Hevy and tap Start.
What you need
- Hevy Pro — the Hevy API key requires it
- Node.js 20+ — https://nodejs.org
- Claude Code (Claude Desktop works too — see below)
Setup (~3 minutes)
1. Get the code and install:
git clone https://github.com/jmilinovich/claude-hevy-coach.git
cd claude-hevy-coach
npm install
2. Add your Hevy API key:
cp .env.example .env
Get your key at https://hevy.com/settings?developer and paste it into .env
so it reads:
HEVY_API_KEY=your-key-here
3. Check it works:
npm run doctor
You want: ✅ Hevy API key works.
4. Start coaching:
claude
When Claude Code asks to use the hevy-coach MCP server, approve it. Then
just say:
coach me
That's it. The first time, Claude interviews you for a couple of minutes (goals, schedule, equipment, injuries, your working weights) and saves your profile. Every time after: it recaps your last session, programs the next one, shows you the card, and — once you say go — writes it to Hevy. Open the Hevy app and tap Start on the 🤖 routine.
Day to day
Things you can say:
- "coach me" / "program my next workout"
- "push day, 45 minutes, easy on the shoulders"
- "how did my last session go?"
- "how's my bench trending?"
- "I tweaked my knee — work around it" (it'll update your profile too)
Your training philosophy lives in COACH.md — edit it any time,
the coach reads it fresh before every session. Your block position, baseline
loads, and prescription history live in state/ (local only, not committed).
Using Claude Desktop instead
Do your first conversation in Claude Code (the setup interview writes your
profile into COACH.md, which Claude Desktop can't edit). After that, Desktop
works great for day-to-day coaching.
Add this to claude_desktop_config.json (Settings → Developer → Edit Config),
replacing both paths with wherever you cloned the repo:
{
"mcpServers": {
"hevy-coach": {
"command": "/ABSOLUTE/PATH/TO/claude-hevy-coach/node_modules/.bin/tsx",
"args": ["/ABSOLUTE/PATH/TO/claude-hevy-coach/src/mcp/server.ts"]
}
}
}
(That's the repo's own tsx binary — Desktop launches servers from outside the
repo folder, so a bare npx tsx wouldn't find it.) Restart Claude Desktop,
then tell it: "You're my strength coach — start with the
get_programming_brief tool." Claude Desktop doesn't read CLAUDE.md, so that
first nudge points it at the brief, which carries the coaching instructions.
One limitation: profile updates ("my gym has no cables now") need COACH.md
edited — do those in Claude Code, or edit the file yourself.
How it works (60 seconds)
- The Hevy API has no scheduler and no deletes — the only way to hand you a
planned workout is to overwrite a routine. So the coach maintains exactly
one routine, marked
🤖(the default marker), rewritten in place each session. COACH.mdis the doctrine — your goals, split, equipment, injuries, and effort rules, in plain English. It's injected into every programming brief.- The engine derives everything from your history — which day is due,
whether it's a deload, what loads to prescribe. The only stored state is a
tiny block seed, your baseline loads, and an append-only log of what was
prescribed (
state/, gitignored). - The learning loop: Hevy records what you did; the coach records what it prescribed. Before each session it diffs the two — held / heavier / lighter / faded / skipped — and adjusts.
- Workout writes are previewed first.
commit_workoutvalidates against your live exercise catalog, shows you the card, and only writes after you approve (creating a custom exercise also waits for your explicit OK). Everything runs locally; the only network calls are toapi.hevyapp.com.
Troubleshooting
| Symptom | Fix |
|---|---|
npm run doctor says key rejected |
Re-copy the key from https://hevy.com/settings?developer; make sure your Hevy Pro subscription is active |
| Claude says it has no hevy-coach tools | Run /mcp in Claude Code and approve the server; make sure npm install ran; restart Claude Code in this folder |
npx tsx not found |
Run npm install in the repo folder first |
| Coach programs movements your gym doesn't have | Tell it — it updates COACH.md and won't do it again |
| On Windows the server won't connect | Edit .mcp.json to {"command": "cmd", "args": ["/c", "npx", "tsx", "src/mcp/server.ts"]} (native Windows needs the cmd /c wrapper; WSL doesn't) |
| You want a fresh start | Delete the state/ folder contents and run git checkout COACH.md (restores the template) — the coach re-interviews you next time |
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。