Your data & AI

This page explains how openkoutsi handles your data, and — if you choose to use the optional AI features — exactly what is sent to a language model (LLM) and where it goes.

How your data is stored

Everything you put into openkoutsi stays on the server your instance runs on:

• Each user has their own private database. Your activities, plans, metrics, and profile are visible only to you — not to other users, and not to an administrator.
• Uploaded FIT files and avatars are stored encrypted on the server.
• You can export your data or delete your account at any time.

Normal use of openkoutsi — uploading rides, syncing from Strava/Wahoo, viewing your metrics and plans — never sends your training data anywhere except between your browser and your own server (and to Strava/Wahoo when you connect them).

Two things you can do that send data elsewhere

Both are deliberate acts, and neither happens on its own:

• Connecting Strava or Wahoo, which is what makes the sync work.
• Creating a personal access token — a long-lived credential you hand to your own tooling so it can read (and, if you grant it, write) your data from outside the browser. Whatever holds that token can pull your data off the instance for as long as the token lives, within the scopes you granted.

A token is not a leak: you create it, you choose its scopes, you see it in a list, and you can revoke it at any moment. But it is the one credential that lets something other than your browser reach your data, so it is worth knowing which tools hold one — and revoking the ones you no longer use.

The AI features are optional

openkoutsi has six features that use a language model:

Feature	What it does
Activity analysis	Written coaching feedback on a single ride.
Daily training status	A short summary of your current form and what to do next.
Goal guidance	Judges how realistic a goal is and how to reach it.
Ask Koutsi	Answers questions you type, by looking up your training data.
AI plan generation	Builds a training plan from your requirements.
AI workout generation	Turns a planned day into a structured interval workout.

Nothing is sent unless you opt in

These features only run when both of these are true:

1. An LLM endpoint has been configured (by you, or instance-wide by your administrator), and
2. You trigger the feature — e.g. you click Analyse on an activity, Generate on a plan, or send a question in Ask Koutsi.

If no LLM is configured, or you never use these actions, no data is ever sent to any model. Every other feature keeps working without it.

How to tell what a model wrote

Koutsi's coaching text is written by a language model, not by a person. Because it arrives in a chat bubble in a coach's voice, that is easy to forget — so every place the text appears carries a short AI-generated note underneath it. You will see it under:

• the AI Analysis on an activity,
• Koutsi Daily Feedback on the dashboard, and
• AI guidance on a goal.

The note is there from the moment the answer starts appearing, not only once it has finished. Training plans are marked differently — an AI tag next to the plan's name in the plan list — because a generated plan is a set of workouts rather than prose.

Treat it as a suggestion

A language model can be confidently wrong. It sees the summary described below and nothing else — not how you slept, not that you were ill last week, not a niggle you have been riding through. Use its feedback as one input alongside your own judgement, and do not treat it as medical or professional advice.

If you are reading your data through the API or your own tooling rather than in the browser, the same disclosure is available there: the fields carrying generated prose ship a companion *_ai_generated flag.

Where your data goes

openkoutsi talks to any OpenAI-compatible chat API. The endpoint is chosen by whoever configures it, and that choice decides whether any data leaves your server:

• A local model (for example Ollama running on the same machine, http://localhost:11434/v1) — the request never leaves your server. Nothing is shared with a third party.
• An external provider (for example a hosted OpenAI-compatible API) — the request is sent over the internet to that provider and is subject to their privacy policy and data-retention terms.

Know your endpoint

Whether your training data leaves your server depends entirely on the endpoint that is configured. If you want your data to stay fully self-hosted, use a local model. If an external provider is configured, treat anything the features send (see below) as shared with that provider.

You can set your own endpoint under Settings → AI / LLM — see Using your own AI model (BYOK). Once you do, only your configuration is used for your account; otherwise the AI features use whatever endpoint your administrator has configured instance-wide.

Choosing which model to use

Your administrator can offer several models to choose from — for example a fast local model alongside one or more hosted providers. When more than one is available, a Model dropdown appears under Profile → Analysis. Pick the one you'd like the AI features to use, or leave it on Default to follow the administrator's choice.

Your model choice can change where your data goes

Each model an administrator sets up points at its own endpoint. Picking a local model keeps your data on the server; picking a hosted provider sends the request (see What is sent to the model below) to that provider, under their privacy and data-retention terms. If you're unsure where a model runs, ask your administrator or choose one you know is local.

What is sent to the model

Only the data relevant to the feature you triggered is included — as a compact text summary, never your raw files. In detail:

Activity analysisDaily training statusAI plan generationAI workout generation

A summary of the one activity you analysed:

• sport, date/time, duration, distance, elevation gain
• average / weighted power, intensity, training load, average / peak heart rate
• the ride's aerobic metrics: efficiency factor, variability index, and the aerobic decoupling figure — or, when decoupling was not measured, the reason why (see Aerobic metrics). These are computed on your server from the ride's data; the underlying per-second streams are never sent.
• your FTP and max heart rate
• your fitness/fatigue/form going into the ride
• the interval breakdown and any personal records set
• the activity's labels and your notes on it (free text you wrote)

A snapshot of your recent training:

• your FTP and max heart rate
• your current fitness, fatigue and form
• your intensity distribution over the last 12 weeks: the share of easy, tempo/threshold and hard work, the shape it makes, and whether your zones changed inside that window (see Intensity distribution). This is computed on your server from your stored zone totals; the underlying per-second streams are never sent.
• the last 28 days of activities (date, sport, duration, training load)
• your active plan's name, dates, and this week's planned workouts
• your active goals (title, target, status)

The requirements for the plan you asked for:

• length, periodization style, intensity preference, training days per week
• your goal/event and any extra description you typed
• your FTP and current fitness
• your intensity distribution over the last 12 weeks, as one line — the shape and the three percentages — so the plan is written against what you have actually been doing

The description of the planned day being expanded:

• workout type and sport
• the planned description, target duration, and target training load
• your FTP

When you let Koutsi look things up

The two lists above describe the normal case: openkoutsi decides in advance what the coach needs, assembles that fixed summary, and sends it once. There is an optional setting — Profile → Analysis → Let Koutsi look things up — that changes how the two coaching features work, and it changes this page's answer along with it. The same setting is what turns on Ask Koutsi, which works this way and no other.

With it on, Koutsi is not handed a summary. It is given a short list of questions it may ask about your own training data, and it decides which ones to ask before writing. That lets it follow a thread: if your form number looks off, it can go and look at the rides behind it and tell you why, instead of repeating the number back to you.

What changes about what is sent

Instead of one fixed summary, several requests are sent, each carrying the answers to the questions asked so far. So:

• The exact contents vary by run. Two days with the same data can send different things, because Koutsi asked different questions.
• More is sent in total. Three to five requests instead of one, and each one repeats what came before. If an external provider is configured, that means more of your training data crossing the internet, more often.
• The boundary is unchanged. Everything under What is never sent below still holds, and the questions Koutsi may ask are a fixed, short list — it cannot invent a new one.

The questions available to it are these, and all of them are read-only and scoped to your own data:

Koutsi can ask	And gets back
How am I doing right now?	Fitness, fatigue, form, recent volume, your FTP and max heart rate
What have I done lately?	Recent activities — date, sport, duration, load
What did I ride on a given day?	Matching activities, searchable by date, sport, session type or name
How did that session actually go?	One activity in detail: intervals, time in zones, notes, RPE, aerobic metrics
Am I on track with my plan?	Active plans, this week's sessions, whether they were done
Will I make my goals?	Your goals, their targets and current progress
How strong am I?	Your power curve and FTP estimate
Have I been training polarized?	Your intensity distribution
Where has my time gone?	Weekly time in each power and heart-rate zone

There is no question that returns your raw ride files, your location data, or anything belonging to another user — those are not on the list, so they cannot be asked for.

You will see what it is doing

While Koutsi is gathering, the card tells you which question it is on — "Koutsi is checking your power curve…" — rather than showing a bare spinner. When the answer starts arriving, the card looks exactly as it always does.

The setting is off unless you turn it on, and turning it back off takes effect on the next run. It also needs a model that supports this way of working: if yours does not, openkoutsi quietly falls back to the fixed summary and the feature behaves exactly as it did before — you do not need to check, and nothing breaks.

Ask Koutsi sends what you type

Ask Koutsi works the same way, with two differences worth knowing about.

The first is that you choose what is sent. Everything else on this page is a summary openkoutsi assembled; a chat question is free text you wrote, and it goes to the model exactly as typed. The same advice as for activity notes applies, more strongly: if an external provider is configured, do not type anything into the box you would not want leaving your server.

The second is that a conversation is kept. Your questions and Koutsi's answers are stored in your own database so you can come back to them, which nothing else here does — the daily card overwrites yesterday's. Earlier turns are replayed to the model on later questions so it can follow the thread, though older ones drop off as a conversation gets long. What Koutsi looked up is not stored or replayed: each question triggers a fresh look at your current data.

You can delete any conversation, and deleting one removes everything in it. Conversations are included in your data export as chat.json and are deleted with your account.

What is never sent

Regardless of which feature you use, openkoutsi does not send:

• your name, email, or login credentials
• your raw FIT files or any GPS / route / location data
• any other user's data
• your stored API key to anywhere other than the endpoint it authenticates against (and keys are held encrypted at rest)

Free-text fields

The features do include free text you wrote — activity notes, goal titles, and plan descriptions — because that context improves the coaching. If an external provider is configured, avoid putting anything sensitive in those fields, or switch to a local model.

Staying in control

• Don't want AI at all? Leave the LLM unconfigured (or don't use the AI actions). Everything else works exactly the same.
• Want it fully private? Point Settings → AI / LLM at a local model, or pick a local model from the Profile → Analysis dropdown if your administrator offers one.
• Changed your mind? Clear the endpoint in your settings; future actions send nothing. You can also export or delete your data at any time.
• Check what else can read your data. Settings → Personal access tokens lists every long-lived credential you have issued, what each one may do, and when it was last used. Revoking one takes effect immediately — see Personal access tokens.