A coach that has actually read your numbers.

Before a question reaches the model, feels.money assembles a compact context block from your own data: a line of profile context, your account balances by type, this month's income and spending with the five categories you spent most in, what your subscriptions cost, your active goals with their progress, and your latest health score with its subscores. Every figure is rounded to whole units. The whole block is capped at around 1,800 characters.

What is deliberately absent matters more than what is present. There are no account numbers, no card masks, no institution credentials, and no raw transaction list. The model cannot see the individual purchase you would rather not discuss; it sees that your Food & Drink total was $742 this month. That constraint is not a limitation we apologise for — it is the design.

The last twelve messages of the conversation ride along with each new question, so follow-ups make sense without the whole history being re-sent. Conversations are stored under your account, and you can delete them — along with everything else — whenever you like. Your data is not used to train AI models, ours or anyone else's. The privacy policy states that plainly and the security page covers how the rest of it is protected.

The coach is instructed to work in a specific order, and the order is the product. Answer the question directly first — nobody wants three paragraphs of preamble before finding out whether they can book the flight. Then the reasoning, in plain language, with the actual figures cited rather than paraphrased. Then two or three concrete next steps, prioritized, specific enough to act on this week.

When maths is involved, it is shown. Comparing two options or projecting a number forward produces a small table you can check line by line, because “you can afford it” is a claim and income minus committed costs minus what you have already spent is a calculation. One is trustworthy; the other is a vibe.

feels.money is not a registered investment adviser, a broker, or a fiduciary, and the coach is not a substitute for one. It is instructed to say so directly whenever a question crosses that line — tax specifics, legal structures, whether to buy a particular security — rather than answering confidently and letting you find out later that it should not have.

It will not guarantee an investment return or predict a market. Where a long-run assumption is used in a projection, it is named as an assumption and the projection is called an estimate. Where a decision depends on things we cannot see — your tax position, a health situation, what you actually want your life to look like — it says that too.

What it will do is the genuinely useful part: explain your own numbers back to you, cost out a decision, compare two options with the arithmetic visible, and translate the jargon. Most financial questions people carry around are not really investment advice questions. They are “can I do this, and what does it cost me?” — and that is a question your own data can answer.

The coach is also instructed never to shame you. Money shame is a remarkably effective way of making someone stop looking at their finances entirely, which is the opposite of useful. Setbacks get framed as data, progress gets acknowledged, and the tone stays the same whether the month went well or badly.

Everything that produces a number — your health score and its eight subscores, budgets and adherence, net worth, cash-flow projections, goal arrival dates, subscription totals, opportunity-cost maths — is pure, tested TypeScript running on our own servers. No model is consulted, so no model can be unavailable. Those figures are always there.

The features that do use a model are built rules-first: a deterministic result is computed before the model is asked for anything, and the model's job is to refine and narrate it. If the backend is slow, down, or returns nonsense, the deterministic result ships anyway. The purchase analyzer, the salary coach, the side-hustle builder, the spending insights and the narrative in your weekly report all behave this way, and each records which engine produced it — rules or ai — so you are never guessing.

A circuit breaker sits in front of the model for exactly this reason. After two consecutive failures it opens for a minute and lets a single probe request through, rather than making every subsequent question burn its full retry budget rediscovering that nobody is answering. The practical difference is between waiting seconds for a rules-based answer and staring at a spinner for minutes.

The chat itself is the one place we deliberately do not fabricate a fallback. If the model cannot be reached, the coach tells you so in plain language and keeps your message saved so you can send it again in a moment. A canned pseudo-answer dressed up as coaching would be a small lie, and this is not a product that can afford small lies.

• “Can I afford this?” — priced against your committed costs, your remaining month, and your goals rather than your account balance alone.
• “Where did the money go?” — a readable account of an expensive month, tied to real categorized spending.
• “What should I do first?” — debt versus savings versus investing, argued from your actual debt load, cushion and cash flow.
• “What does this word mean?” — APR, vesting, expense ratio, index fund, explained once and without condescension.
• “How is this goal going?” — the arrival date recalculated from your real contribution pace, not the one you hoped for when you set it.

And the ones to take elsewhere: your specific tax return, whether to exercise your options, how to structure a business, whether a particular stock is a good buy. The coach will tell you the same thing.