Pulsai Try Pulsai →

Garmin + AI · 2026 guide

Garmin MCP: open-source vs hosted

There's a growing set of Garmin MCP servers that let an AI assistant read your Garmin Connect data. Most are open-source and self-hosted; Pulsai is hosted. Here's an honest look at both — so you pick what fits, not what we sell.

What is a Garmin MCP server?

MCP (Model Context Protocol) lets AI assistants call external tools. A Garmin MCP server exposes your Garmin Connect data — activities, sleep, HRV, training load, power, and more — as tools your assistant can query in plain language: “How did I sleep before my best 10K this year?”

The open-source projects run locally on your machine (you clone a repo, install dependencies, run it, and wire it into Claude Desktop). A hosted server like Pulsai runs in the cloud and gives you a private URL — no install.

Quick comparison

 Open-source (self-hosted)Pulsai (hosted)
SetupTerminal: clone, install (uv/npm/Docker), runPaste one URL — no terminal
AI clientsMostly Claude Desktop onlyClaude, ChatGPT & Gemini
Runs onYour computer (must stay on)Hosted — works on web & mobile
CredentialsOften plaintext env vars / long-lived tokens on diskPassword never stored; tokens encrypted at rest
Writes workouts backUsually read-onlyReads live & writes workouts
CostFree (your time + your machine)Free tier + paid Pro
Control / privacyFull — nothing leaves your machine*You trust the host (encrypted, deletable)

*Once you ask your AI about your data, the values you request go to that AI provider (OpenAI/Anthropic/Google) — that's true for any option, self-hosted or hosted.

The open-source options

Taxuspt/garmin_mcp — the most complete

110+ tools covering ~90% of the python-garminconnect library, including advanced cycling analytics (power duration curve, VAM climb detection, aerobic decoupling, W/kg) and FIT/GPX downloads. Excellent if you're comfortable with the terminal.

Needs: clone the repo, install with uv, run locally, wire into Claude Desktop. github.com/Taxuspt/garmin_mcp

Full disclosure: Pulsai is built on this same open-source engine — hosted, multi-client, and multi-tenant. Credit where it's due.

Nicolasvegam/garmin-connect-mcp

A clean TypeScript server with 61 tools. Per its docs, it takes your Garmin email and password as environment variables — convenient, but plaintext credentials on disk is a pattern a hosted service can avoid. Repo

brunosantos/garmin-workouts-mcp

Focused on workout creation. The author is refreshingly honest that it relies on long-lived tokens (a security trade-off vs OAuth best practice), and notes that Garmin's official API program is enterprise-oriented — an individual open-source developer application was rejected. It's the clearest illustration of why a compliant, hosted layer is hard to DIY. Repo

Others — Jack-Abyss/claude-garmin, eddmann, st3v, benniblau

Similar variants: terminal install (PowerShell, Docker, or a Python venv), mostly Claude Desktop only, and single-user. Great for tinkerers; more friction if you just want answers.

When to choose open-source

When to choose hosted (Pulsai)

Skip the setup — talk to your Garmin data in 2 minutes

Pulsai is a hosted Garmin MCP for Claude, ChatGPT & Gemini. No terminal, no install.

Connect your Garmin →

FAQ

Is there a Garmin MCP that works without the terminal?

Yes — a hosted Garmin MCP like Pulsai gives you a private URL you paste into your AI client. No cloning, no installs, no server to keep running.

Which Garmin MCP works with ChatGPT and Gemini?

Most open-source projects target Claude Desktop only. Pulsai's hosted Streamable HTTP endpoint works with Claude, ChatGPT and Gemini from one URL.

Are open-source Garmin MCP servers safe?

They run locally, which some prefer — but several store credentials as plaintext env vars or long-lived tokens on disk. A hosted option encrypts tokens at rest and never stores your password, at the cost of trusting the host (which is why Pulsai lets you delete everything anytime).

Repo names and details are referenced from each project's public GitHub documentation and are the property of their respective authors. This comparison aims to be fair — open-source and hosted each fit different people.