EKIDEN.AI v3

Start

EKIDEN.AI v3 is an analysis layer that reads the FIT file from your watch and works out what the workout actually was: 800 m × 10, where you slowed down, whether you got faster than last time. It returns numbers and judgments; the AI you connect (Claude or ChatGPT) puts them into words.

• Rules count: reps, distances, times and paces. The same record always gives the same answer.
• Jev judges: whether the run was planned and what it aimed at, chosen from fixed options with a probability.
• The athlete decides the rest: only what only they can know, asked as a multiple-choice question.

Two ways to use it

	What you need	Your data
MCP (Claude Desktop extension)	Claude Desktop	Stays on your computer
CLI	Node.js 20 or later	Stays on your computer

A remote version (add it to Claude or ChatGPT by entering a URL) is in preparation.

Testing in progress

Downloads are at v3.ekiden.ai/cli/ and are for the testing team only (password). After the public release, the CLI will be published on npm.

MCP (Claude Desktop extension)

1. Download ekiden-v3-local.mcpb.
2. Double-click it. Claude Desktop shows an install screen; click "Install".
3. Choose your FIT folder: the folder with .fit files exported from your watch, or .zip exports from Garmin Connect or COROS (subfolders are read too).
4. Optionally, fill in the following.

Setting	Example	Used for
Recent race	marathon:2:58:40, 10k:38:42	Pace zones and predicted times
Max and resting heart rate	190, 48	Heart rate zones
Language	ja / en	Result text

5. Ask Claude.

List my interval sessions from last month Show me the July 15 workout segment by segment Am I getting faster at 1000 m × 5?

The available tools are listed under Tools. Supported: macOS and Windows.

Tools

All tools are read-only. Results come back as a one-sentence summary plus JSON, with no location data (no GPS coordinates). Large results are paged with offset.

Tool	What it does
list_activities	List workouts, filtered by period, sport and shape
get_activity	One workout in detail: summary, laps and segments
get_splits	First and second half, per km, or per segment
compare_activities	Compare 2 to 5 workouts
get_activity_window	Zoom into part of a workout, such as "the last 5 minutes" or "the 3rd rep"
get_training_summary	Training volume by week or month
get_athlete_profile	VDOT and predicted times for each distance
get_athlete_zones	Pace zones and heart rate zones
get_menu_history	How the same session has progressed, and whether you are getting faster

CLI

You need Node.js 20 or later (check with node --version). In the folder where you downloaded ekiden.mjs:

node ekiden.mjs --help
node ekiden.mjs analyze ~/fit --race marathon:2:58:40
node ekiden.mjs list-activities --local ~/fit
node ekiden.mjs get-splits --activityId <ID> --mode segments --local ~/fit
node ekiden.mjs compare-activities --activityIds <ID1>,<ID2> --local ~/fit

Replace ~/fit with your FIT folder. Add --json for JSON output and --lang en for English. node ekiden.mjs features lists the available features.

Pass thresholds with --vdot 54, --race <mile|5k|10k|half|marathon>:<time>, or --t 4:00 --i 3:40 --m 4:14 --e 5:06. Heart rate zones use --max-hr 190 --resting-hr 48.

After the public release, the CLI will be published on npm (so that it can be run with npx).

Training Catalog

Sessions are identified using the Training Catalog by Arata Fujiwara. Arata built it from his own experience and what he learned from books and other sources; it is his work.

	Count
Sessions	191 (11 categories)
Session types	8: recovery, easy, threshold, vo2max, marathon, canova, long, workout
Training plan styles	Daniels, Canova

When the catalog has no exact match, the session is named from the structure the rules counted (for example, 800 m × 10) combined with the aim Jev chose (for example, vo2max).

Your data

• Read only: it never writes to your training records.
• No location data: results passed to the AI contain no GPS coordinates.
• Local versions stay on your computer: the MCP extension and the CLI read your FIT files on your own machine.
• No conversation: v3 only returns numbers and judgments. The AI you connect writes the text.