Fartlek

A coach's morning report from your Garmin data, for any LLM via MCP.

Every other Garmin MCP server hands the LLM a filing cabinet of raw JSON — one night of sleep is ~52K tokens, one activity stream ~155K. The model can't read it, so it skims and improvises. Fartlek does the synthesis server-side: computed sports-science metrics (CTL/ATL/TSB, ACWR, monotony, calibrated training load), personal baselines with significance floors, safety alerts — delivered as compact, verdict-first reports the model can actually reason about.

The token contract (v0.1): calling every tool in the catalog once, at default arguments, costs under 9K tokens — a sixth of one raw Garmin sleep payload. Excluding the garmin_raw escape hatch, the whole synthesis surface sums to under 4K. Hard caps are enforced per response by the renderer, with disclosed truncation.

Status: v0.1 (Phase 1). 8 synthesis tools; the trend suite (weekly review, multi-week load, fitness/race outlook, recovery audit) ships with v0.2. Design: docs/DESIGN.md · plan: ROADMAP.md · contributors: docs/HANDOFF.md.

The tools

First call on a fresh install runs the cold start automatically (~30 API calls, ≈1 minute): 180 days of history, warm CTL/ATL from day 0, then background sleep/HRV backfill.

Related MCP server: claude-garmin

Quickstart

Install with either:

# uv (recommended) — runs without cloning
uvx fartlek-mcp

# or pipx
pipx install fartlek-mcp

Or clone and run from source (requires Python ≥ 3.12 and uv):

git clone https://github.com/matisdsp/fartlek && cd fartlek
uv sync

# One-time Garmin login (email/password + MFA if enabled).
# Credentials are never stored; OAuth tokens go to ~/.fartlek/tokens/.
uv run fartlek auth

# Optional but recommended: warm the local store now instead of on first use
uv run fartlek sync --nights 60

uv run fartlek doctor   # check everything is healthy

Then point your MCP client at the server.

Any MCP-compatible client works — the server speaks standard JSON-RPC over stdio, so it is client-agnostic. The snippets below are just the per-client config formats; Claude Desktop, Claude Code, Cursor, Continue, Cline, Windsurf, Zed, VS Code (Copilot Chat), and Gemini CLI all work. The universal invocation is uvx fartlek-mcp.

Claude Code — from this directory, .mcp.json is picked up automatically. From anywhere else:

claude mcp add fartlek -- uvx fartlek-mcp

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "fartlek": {
      "command": "uvx",
      "args": ["fartlek-mcp"]
    }
  }
}

Cursor — .cursor/mcp.json, same command/args block as above.

Continue / Cline / Windsurf / Zed — same pattern: wherever the client keeps its MCP server list, add a fartlek entry with command: "uvx", args: ["fartlek-mcp"]. Most editors adopt the Claude Desktop format verbatim.

Any other stdio MCP client — invoke the server binary directly:

fartlek-mcp          # speaks JSON-RPC over stdin/stdout

Ask things like "can I go hard today?", "analyze my last run", "how did I sleep this week?" — and tell it how sessions felt: your reported RPE and illness notes gate the readiness verdict.

Docker

Build and run locally:

docker build -t fartlek-mcp .
# Tokens and store are persisted in ./fartlek-data on the host
mkdir -p fartlek-data
docker run -i --rm -v "$PWD/fartlek-data:/data" fartlek-mcp

For fartlek auth, run it interactively once to populate the volume, then use the image as the MCP server:

docker run -it --rm -v "$PWD/fartlek-data:/data" --entrypoint fartlek fartlek-mcp auth

CLI

Environment: GARMINTOKENS overrides the token location, FARTLEK_HOME the data directory (default ~/.fartlek).

Releasing to PyPI (maintainers)

Releases are published via trusted publishing (OIDC) — no API tokens anywhere.

1. On pypi.org → Account settings → Publishing → add a GitHub publisher:

   • PyPI project name: fartlek-mcp · owner: matisdsp · repo: fartlek

   • Workflow: release.yml · environment: pypi

2. Bump version in pyproject.toml, commit, then tag and push:

git tag v0.1.0 && git push origin v0.1.0

The release workflow builds, runs tests, and uploads to PyPI. A published version can't be overwritten — to fix a mistake, bump to the next patch (0.1.1).

Privacy

Local-first: stdio transport, your credentials and health data never leave your machine. The server only talks to Garmin's API with your own tokens, sequentially and rate-limited. fartlek export gives you everything; fartlek reset removes everything.

How Fartlek reaches your data — read this before connecting an account

Garmin has an official developer programme, and Fartlek is not part of it. Like every other open-source Garmin client, it signs in with your own credentials and reads the same endpoints the Garmin Connect apps use. Those endpoints are not published for third-party use, and Garmin's Terms of Use list, among examples of prohibited conduct, "using any process, whether automated or manual, that accesses, copies, or scrapes content from the Site through any means not purposely made available through the Site."

What that means in practice:

• Your account is yours to risk. Garmin can rate-limit, block, or suspend accounts for automated access. Fartlek is deliberately polite — sequential calls, backoff on 429, and Garmin is contacted only by the sync process, never per question — but politeness is not permission.

• It can break without warning. Garmin changed its login in March 2026 and broke every third-party client for weeks. This will happen again.

• Read-only, by decision. Fartlek never writes to Garmin: no workouts pushed to your watch, no training plans, no edits to your activities. The two tools that write (garmin_log, garmin_set_profile) write to the local SQLite store and nothing else. Pushing structured workouts was specified and then dropped — see docs/DESIGN.md §2.4. Garmin's Training API is the sanctioned route for that, and it requires a cloud-to-cloud integration, which would mean your data leaving your machine.

• Nothing is redistributed. Your data stays on your disk. Fartlek's own responses are derived from it and are shown only to you and the LLM client you chose.

If that trade-off is not one you want to make, do not connect an account. This is stated here rather than buried, because it is the kind of thing you should decide before installing, not discover afterwards.

License & trademark

Apache 2.0. Fartlek is an independent open-source project, not affiliated with, endorsed by, or sponsored by Garmin Ltd. "Garmin" is used only to describe compatibility with Garmin Connect data.