TrainBud

Talk to your training data.

TrainBud is an open-source MCP server that connects your Garmin Connect fitness data to Claude, Cursor, and other AI assistants. Ask about workouts, sleep, heart rate, recovery, and body composition in plain English — privately, on your machine.

Disclaimer: TrainBud is an unofficial community project. It is not affiliated with, endorsed by, or sponsored by Garmin Ltd. Garmin Connect is a trademark of Garmin Ltd.

Try it

Once connected to your MCP client, ask things like:

• "What did I do today?"

• "How's my sleep been this week?"

• "Am I recovered enough to train hard tomorrow?"

• "Is my resting heart rate trending down?"

See examples/prompts.md for more ideas.

Related MCP server: Garmin Connect MCP Server

Why TrainBud

• Private — credentials stay in your local .env; data is cached on your machine

• Local-first — SQLite cache, session tokens in .trainbud/

• Works everywhere — Windows, macOS, Linux (Node.js 22.12+)

• Any MCP client — Claude Desktop, Cursor, and other stdio-compatible clients

• Smart fetching — batched API calls and automatic re-auth when sessions expire

Quick start

npx trainbud setup

The setup wizard walks you through credentials, authentication, and connecting Cursor or Claude Desktop — no MCP config editing required. Then restart your MCP client and ask it what you did today.

To keep trainbud on your PATH instead of typing npx every time:

npm install -g trainbud
trainbud setup

Needs Node 22.12+. Full walkthrough: QUICKSTART.md

git clone https://github.com/Zsadigzade/trainbud.git
cd trainbud
npm install
npm run build
npm link          # puts `trainbud` on your PATH; undo with `npm unlink -g trainbud`
trainbud setup

Without npm link, every trainbud <command> in this README is node dist/index.js <command> run from the repo root. If dist/ does not exist yet, run npm run build first.

Claude Code plugin (recommended)

Install as a Claude Code plugin — skills and MCP server in one step:

/plugin marketplace add Zsadigzade/trainbud
/plugin install trainbud@trainbud

Set credentials, then restart Claude Code:

export GARMIN_EMAIL="your@email.com"
export GARMIN_PASSWORD="yourpassword"

Plugin files live in plugin/. See plugin/README.md.

Claude Code skills (in-repo)

This repo also ships project skills in .claude/skills/ for development without installing the plugin:

Open the repo in Claude Code (claude in this directory) — skills load automatically.

To use skills in every project without the plugin, copy them to ~/.claude/skills/.

After setup, restart your MCP client and try /trainbud with "What did I do today?"

Dashboard

trainbud serve hosts a dashboard at /dashboard. It is phone-first, because the pairing flow is: you are standing next to the watch holding a phone when you approve a code.

It shows what stands out today against your own baselines, this week against last week, and resting heart rate and sleep plotted against your own 30-day median — all read from the local history store, so it paints instantly and works even when your Connect session has expired. A break in a line is a day with no measurement, not a zero.

It is also where you tell TrainBud who it is talking to:

Usage. TrainBud runs on your own AI provider key, so every question and every daily insight is charged to you. The dashboard shows the tokens and cost per call, the month to date, and a 30-day chart. A model this build has no published price for is recorded with its cost left unknown rather than as zero — a call priced at zero would make a cap that can never trip.

Nothing on this page leaves your machine. There is no endpoint to send it to.

Garmin watch widget (Connect IQ)

View recovery, sleep, activity, stress, and VO2 max on your Garmin watch via a Connect IQ widget in ciq/.

Requires: trainbud serve running + HTTPS tunnel (same setup as web AI).

1. Start the server and tunnel:

   trainbud serve
   cloudflared tunnel --url http://127.0.0.1:3847

2. Build and sideload the widget — see ciq/README.md

3. In Garmin Connect Mobile → Connect IQ → TrainBud settings, set:

   • Server URL — your tunnel URL (e.g. https://abc.trycloudflare.com)

4. Open the widget on your watch — it shows a pairing code. Approve it in the dashboard (/dashboard?token=YOUR_API_KEY) to complete setup. The dashboard swaps that token for a session cookie and drops it from the URL, so the address bar is safe to screenshot afterwards.

The glance shows recovery and sleep, or the top finding in a few characters ("RHR +4 bpm"), from the last cached summary, so it renders without waiting on the network; once that is more than two hours old it says so. Open it and tap or swipe to cycle through the cards you left switched on in the dashboard. The watch calls GET /api/watch — a compact JSON summary, not the full MCP protocol.

On the Today card, START (or tap) opens a finding to the rule it fired on — "Each of the last 3 days was at least 3 bpm and 2 deviations above your 28-day median of 48 bpm" — and from there it can be muted for three days (POST /api/mute). The Sleep card names what moved last night against your own nights ("Deep 40m (1h24m)"). On the Forerunner 55, whose widget has 64 KB, those two are compiled out so the rest still fits.

Detector rules

Every finding fires on a rule you can see and move. The dashboard's Detector rules section holds the bars — resting HR days/bpm/deviations, sleep-debt hours, HRV drop, the training-load ratios, and "several signals together" (resting HR up, HRV down and sleep stress up at once) — bounded so no setting makes every day, or no day, a finding. Every surface shows the rule a finding fired on, with the numbers that were in force.

Voice (/voice)

Hold a button on your phone, ask a question out loud, hear the answer read back. Open https://YOUR-HOST/voice?token=YOUR_API_KEY once; it trades the key for a cookie and redirects to a clean URL.

• Android / Chrome — speech recognition happens in the browser. Free, no key.

• iPhone / Safari — records audio and transcribes it server-side via Groq (whisper-large-v3-turbo, about $0.04 per hour of audio). Add the key on the dashboard.

• Both — answers are spoken back, and Read me my day / Read today's insight cost nothing at all: they are composed in code from data already on hand, not generated.

Not a watch feature and it cannot be one: Connect IQ exposes no microphone or speaker to a widget. Not hands-free either — mobile browsers suspend the microphone when the tab is backgrounded, so the phone has to be awake with the page in front of you.

Full detail: docs/VOICE.md

Always on (/voice and the watch, without babysitting a terminal)

trainbud serve in a terminal lasts until the terminal closes. To have the server and the tunnel come back by themselves after a reboot:

.\scripts\install-always-on.ps1 -Hostname trainbud.example.com

That registers a scheduled task for the server and a watchdog that checks both halves every five minutes — because a tunnel that answers is not a server that answers: a tunnel with nothing behind it serves its own error page under a 200, so the link looks alive while the watch reports HTTP -400. systemd, launchd and Docker Compose recipes are in the guide.

Full detail, including Cloudflare named tunnels: docs/ALWAYS-ON.md

Connect to Claude Desktop

Edit claude_desktop_config.json:

Windows: %APPDATA%\Claude\claude_desktop_config.json
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "trainbud": {
      "command": "node",
      "args": ["C:/path/to/trainbud/dist/index.js", "start"],
      "env": {
        "GARMIN_EMAIL": "your@email.com",
        "GARMIN_PASSWORD": "yourpassword"
      }
    }
  }
}

With trainbud on your PATH (npm install -g trainbud, or npm link from a clone), point the client at the command instead of a path:

{
  "mcpServers": {
    "trainbud": {
      "command": "trainbud",
      "args": ["start"]
    }
  }
}

Restart your MCP client, then start asking questions.

Tools

CLI

trainbud setup          # Interactive first-time setup (recommended)
trainbud serve          # Remote HTTP MCP for web AI (claude.ai, ChatGPT)
trainbud check          # Live diagnostics against all tools
trainbud doctor         # What the watch would see: public URL, AI key, history depth
trainbud backfill       # Pull Garmin history into the local store (resumable)
trainbud findings       # What stands out against your own baselines, and the rule each fired on
trainbud rederive       # Rebuild sleep measurements (deep, REM, awakenings) from the local archive — no Garmin requests
trainbud start          # Start the MCP server (stdio)
trainbud auth           # Force re-authentication
trainbud cache clear    # Clear cached data
trainbud devices        # List paired watches
trainbud devices revoke <id>   # Take one watch's access away
trainbud rotate status  # Which secrets are set, and whether their two homes agree
trainbud rotate api-key # Replace TRAINBUD_API_KEY, and say what that invalidates
trainbud rotate ai-key  # Replace the AI key in BOTH .env and app.db
trainbud status         # Show session and cache status
trainbud --version      # Print version

rotate ai-key writes both stores on purpose. The AI key lives in .env and in app.db, the database copy is the one actually read, and hand-editing only the file leaves a rotated key that never took effect.

Every one of these also works as npx trainbud <command> without installing anything.

Troubleshooting: trainbud: command not found

Either use npx trainbud <command>, or install it globally with npm install -g trainbud. Running from a clone instead? Run npm link once from the repo root, or call the built entry point directly with node dist/index.js doctor (after npm run build).

Configuration

Security & privacy

• Credentials live only in your local .env file — never sent to a third party

• Session tokens in .trainbud/session.json are as sensitive as a password

• Tool errors are sanitized before reaching the AI client

• Uses the unofficial garmin-connect npm package (not Garmin's enterprise OAuth API)

• MFA is not supported by the underlying library — disable MFA or use an app-specific password

• The server binds 127.0.0.1 by default. It is only reachable from the internet if you point a tunnel at it, and every route except /health needs the API key

• The dashboard takes the key once, on /dashboard?token=…, then trades it for an HttpOnly session cookie and redirects to a clean URL — so the key does not sit in your address bar, your history, or a screenshot

• A paired watch holds a token scoped to that watch, minted at pairing and stored on the server as a SHA-256 hash. trainbud devices lists them, trainbud devices revoke <id> takes one away — without logging out the dashboard, /mcp, or your other watches. A watch paired before 0.5.2 holds the API key itself; re-pair it to swap that for a scoped token

• Every response carries Content-Security-Policy, X-Content-Type-Options, X-Frame-Options and Referrer-Policy, including the 401s. HSTS is sent only on a request that actually arrived over TLS, so the loopback dashboard stays reachable

What TrainBud is not

• Not a hosted service. There is no TrainBud account and no TrainBud server. You run it, on your machine, against your own Garmin credentials

• Not an official Garmin integration. It drives an unofficial library against the Connect web API. Garmin can change that API without notice, and does

• Not MFA-compatible. If your Connect account has MFA on, this will not log in

• Not free to ask. The AI features run on your own Anthropic key and are billed to you. The dashboard meters every call and can refuse past a cap you set

Troubleshooting

Docker

A published image, built and exercised in CI on every push so it is not a promise that quietly rots:

docker pull ghcr.io/zsadigzade/trainbud:latest

Or build it yourself from this repository:

docker build -t trainbud .

One thing matters more than the rest: mount a volume at /app/.trainbud. That is where the SQLite history, the cached session and the log live. Without it, every container restart throws away your downloaded history and re-authenticates from scratch.

As an MCP server for a desktop client

The default entrypoint speaks MCP over stdio, so the client runs the container:

{
  "mcpServers": {
    "trainbud": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "GARMIN_EMAIL", "-e", "GARMIN_PASSWORD",
        "-v", "trainbud-data:/app/.trainbud",
        "trainbud"
      ]
    }
  }
}

-i is required — without it the container has no stdin and the client sees a server that connects and immediately goes quiet. The two -e flags with no value pass the variables through from your own environment, so your credentials stay out of the config file.

As an HTTP server, for the watch app or a web client

docker run --rm -p 3847:3847 \
  -e GARMIN_EMAIL -e GARMIN_PASSWORD \
  -e TRAINBUD_API_KEY \
  -e TRAINBUD_HOST=0.0.0.0 \
  -v trainbud-data:/app/.trainbud \
  --entrypoint node trainbud dist/index.js serve

TRAINBUD_HOST=0.0.0.0 is needed inside a container: the server binds 127.0.0.1 by default, which is the right default on a laptop and unreachable from outside a container.

Filling the history

trainbud setup is interactive and assumes a terminal, so with Docker do the one-off backfill directly:

docker run --rm -e GARMIN_EMAIL -e GARMIN_PASSWORD \
  -v trainbud-data:/app/.trainbud \
  --entrypoint node trainbud dist/index.js backfill

The same pattern runs doctor, check and devices.

Development

npm install
npm run build
npm test          # the whole suite, via the Node test runner
npm run lint
npm run dev       # Start with auto-reload

Use .nvmrc with nvm/fnm for Node 22. npm run test:coverage reports coverage through Node's own test runner, and npm run test:watch reruns on change.

See CONTRIBUTING.md and docs/VAULT.md for architecture and design notes (Obsidian vault, outside this repo).

Roadmap

• VO2 max trends

• Stress levels

• Training insights

• Workout comparison

• Docker image — published to ghcr.io/zsadigzade/trainbud

License

MIT — see LICENSE.

Garmin Connect is a trademark of Garmin Ltd. This project is not affiliated with Garmin Ltd.