🏃 Slipstream

Chat with your Garmin training data in Claude and ChatGPT - free, self-hosted, and private.

Slipstream is a small, self-installing pipeline that pulls your Garmin activities on a schedule and serves them to AI assistants as a custom MCP connector. Ask things like "how far did I run this month?" or "given my recent fitness, is a 14 km run too much today?" - from your phone, the web, or anywhere you use Claude or ChatGPT.

A free alternative to Strava's paid MCP. Strava's official AI connector requires a paid subscription. If you wear a Garmin and want the same conversational access to your own training - without paying - this is for you. Your data, your cloud, no middleman.

---

What you can ask

Once connected, just talk to your data:

• "How much did I run and cycle in the last 30 days?"
• "What's my longest run, and my fastest average pace?"
• "Break down my training by sport this month."
• "I want to run a bit longer than usual - is 12-14 km reasonable given my recent fitness?"
• "Am I training more or less than last month?"

Want more ideas? docs/PROMPTS.md has dozens of example prompts and use cases, from race planning and readiness to recovery, heart-rate and fueling angles, and a "be my coach" mode.

How it works

   Garmin Connect
        │  (python-garminconnect, token auth - no password stored)
        ▼
  GitHub Actions  ──── every 6h + on-demand ────►  data/activities.csv
   (your private repo)                              (committed automatically)
        │
        │  read-only, on demand
        ▼
  Cloudflare Worker  ◄──── MCP (Streamable HTTP) ────  Claude  ·  ChatGPT
   (free, always-on)                                   (phone, web, agents)

• Fetch - a scheduled GitHub Action pulls your recent activities and commits a tidy data/activities.csv to your private repo. No server to run.
• Serve - a tiny Cloudflare Worker (free tier) reads that file on demand and exposes it over MCP at an unguessable URL. Summary data only - no GPS or home-location is ever exposed.
• Ask - you add the Worker URL as a custom connector in Claude and/or ChatGPT. It works everywhere those apps do, including your phone.

Everything runs on free tiers and your own accounts. There is no shared backend.

Quick start

You'll need (all free): a Garmin account, a GitHub account, a Cloudflare account (free; the installer walks you through signup if you don't have one), and Claude and/or ChatGPT. Plus python3, node, git, and the GitHub CLI (gh) on your machine.

1. Create your private copy and start the installer - one command (needs the GitHub CLI):

   gh repo create slipstream --template yhecht/slipstream --private --clone && cd slipstream && ./setup.sh

   No gh? Click “Use this template” on GitHub (make it Private), clone it, then run ./setup.sh.
2. Follow the prompts. The installer handles Garmin login (2FA), the first data fetch, and the Cloudflare deploy - opening browser pages when needed - then prints (and copies to your clipboard) your connector URL. It's safe to re-run; it resumes where it left off.
3. Paste the URL into Claude and/or ChatGPT (it opens the page and shows the exact clicks; full guide in docs/CONNECT.md). Ask it something. Done.

That's it. From then on your data refreshes automatically every few hours.

Connecting the assistants

The installer prints your URL and the steps. For the full click-by-click walk-through, including the easy-to-miss bits, see docs/CONNECT.md. In short:

Claude (free plan; claude.ai, syncs to the mobile app)

Settings → Connectors → Add custom connector → paste the URL (leave the OAuth fields blank) → Add → Connect. Then enable it in your chat using the + / tools control next to the message box.

ChatGPT (paid plan; on the web)

Settings → Apps & Connectors → Advanced → enable Developer mode → Create → paste the URL → set Authentication to No authentication (it defaults to OAuth, which won't work) → tick the consent box → Create. Then, in a chat, + → More → Slipstream to switch it on.

Always paste the URL, never type it. It ends in a long random secret; one wrong character and it won't connect.

Privacy & security

• Your data stays in your own private GitHub repo. There's no shared server.
• No GPS / home location is exposed by the connector - only activity summaries (distance, duration, heart rate, sport, etc.).
• The connector lives at an unguessable URL (a long random path). Treat that URL like a password - it's the only key to your data.
• Your Garmin password is never stored - login mints a refreshable session token that lives only as encrypted secrets.
• For the simplest setup the Worker reads your data repo using your existing GitHub login. Want least-privilege instead? Swap it for a read-only token (see Hardening below).

Hardening (optional)

Replace the data-read credential with a read-only, single-repo token:

1. Create a fine-grained token at https://github.com/settings/personal-access-tokens/new
   • Repository access: Only select repositories -> your slipstream repo
   • Permissions -> Repository -> Contents -> Read-only
2. Store it on the Worker:

   cd worker && printf "%s" "<your-token>" | npx wrangler secret put GITHUB_TOKEN

Now, even in a worst case, the connector can only ever read your activity data.

Keeping data fresh

The GitHub Action runs every 6 hours automatically. To refresh on demand (e.g. right after a workout), trigger it from anywhere:

gh workflow run refresh.yml

You can even wire this to a one-tap phone shortcut - see docs/PHONE_TRIGGER.md.

Troubleshooting and FAQ

Hit a snag? Every common issue across Garmin, GitHub, Cloudflare, Claude, and ChatGPT (plus things outside this project's control) is covered in docs/TROUBLESHOOTING.md. Broader questions (cost, privacy, how it compares to Strava's MCP, supported devices) are answered in docs/FAQ.md.

Documentation

• Connecting Claude & ChatGPT - click-by-click setup for both
• Architecture - how it is built and the data flow
• Troubleshooting - every common issue, by area
• FAQ - cost, privacy, comparisons, limits
• Example prompts - what to ask, by use case
• Phone trigger - one-tap refresh from your phone
• Security policy - posture, reporting, and known limitations
• Contributing - how to propose changes

How it's built

A deeper tour of the design is in docs/ARCHITECTURE.md. In short:

Piece	Tech
Fetch pipeline	Python · python-garminconnect
Scheduler	GitHub Actions
Connector	Cloudflare Workers · agents / MCP SDK · TypeScript
Protocol	MCP over Streamable HTTP

Roadmap

• ✅ Garmin - runs, rides, swims, and everything your watch records.
• ⏳ Wahoo (bike power) and Strava backfill - planned; the architecture already supports multiple sources with de-duplication.

License

MIT © Yannique Hecht

---

_{Slipstream is an independent, open-source project. "Garmin", "Claude", and "ChatGPT" are trademarks of their respective owners; this project is not affiliated with or endorsed by any of them.}