Why not just use ChatGPT or Claude?

The fair version

It works, right up until it matters

This is a reasonable thing to ask, and the honest answer starts by conceding the point: typing "bench 3x8 at 185" into a chat works. The model understands it, it can talk about it sensibly, and for the first few weeks the experience is barely different.

The difference doesn't show up in the logging. It shows up in the asking — the first time you want a number back that you can act on, and need to know whether it's real.

The first problem

A conversation is not built to keep things

Chat memory is designed to carry what's salient about you between sessions — how you like answers, what you're working on, the shape of your situation. That's a genuinely useful job, and it is not the same job as holding several hundred structured records where every individual number has to survive intact and stay retrievable.

Nothing is being done badly here; it's a design aimed at a different target. A training log has an unusual requirement — it is worth almost nothing in the moment and everything in aggregate, two years later, at a level of detail nobody would call salient at the time.

The serious problem

It will answer even when it doesn't know

Ask a model what you squatted in March and, with nothing to look the answer up in, it will produce the most plausible continuation — a sensible-looking number, in the right range, delivered with the same confidence as a real one.

That's the part that should worry you, because it fails silently. A spreadsheet with a missing row shows you a gap. A conversation with a missing session hands you a number that looks exactly like the true ones, and you plan your next block around it.

Workout Chat can misread a sentence on the way in — any parser can, and you correct it in the next message. What it cannot do is invent a session that was never logged, because answers about your history are read back from stored rows rather than recalled.

The quiet problem

You can't compute anything over prose

An estimated one-rep max needs a weight and a rep count as numbers. A volume total needs sets times reps times load, per exercise, across months. A trend line needs a fitted series of dated points. None of that can be derived from a conversation, however well the conversation went.

• Progression charts, on the right metric per movement
• Estimated 1RM, from sets of twelve reps or fewer
• Volume over time, compared against the equal window before it
• Records that update themselves, with recent ones flagged
• Injuries linked to the session and movement that caused them

These aren't features layered on top of the chat. They're what having a schema buys you, and they're unavailable at any price without one.

The one that lasts

You can take a database with you

Your training history should outlive your choice of app, and probably your choice of AI provider. Workout Chat exports to CSV or JSON in one click, or reads out over the REST API — typed fields per exercise: sets, reps, the weight you entered, the weight in kilos, duration, distance, dates, notes, and your weigh-ins alongside.

A chat transcript also exports, but what you get is a record of a conversation. Turning years of that back into something you can chart is a data-recovery project, and you'd be doing it with no way to check the result against anything.

The bit people miss

This isn't ChatGPT or Workout Chat

Workout Chat has its own chat, and it also runs a remote MCP server. Connect that and you carry on logging from inside ChatGPT, Claude, or Cursor — same window, same sentence, same habit.

What changes is where the sentence lands. Instead of dissolving into a conversation, it becomes a row you can chart, query, correct, and export. The assistant stays your interface; it just stops being your filing system.