Skip to content

Draft: not yet reviewed by a clinician

Use this page to prepare questions for your care team. It does not replace their advice. Your own targets and doses come from your nephrologist, diabetes team and renal dietitian.

Optional AI

Off until you choose it

AI features are off unless your admin switches them on, and then off for you until you opt in.

The app works fully without AI. AI is an extra layer that can re-rank meal ideas, turn a sentence such as "2 eggs, toast with butter, tea" into food searches, or read the numbers from a photo of a label (design note 04). The app's own rules check every AI answer before you see it.

How the setup fits together

Three choices make up every AI setup (Settings → AI ideas starts with the same three questions):

  1. Which AI answers. Your admin chooses it: a model running at home, a company's AI over the internet (such as OpenAI or OpenRouter), or a home "Hermes" agent. Your admin may also let you add your own key.
  2. What it may do. Your admin switches AI on, and then you opt in. AI can suggest meals, turn a typed meal into food searches, and read a label photo. If your admin also switches on Ask, it can answer a short question about your food and numbers. It cannot give doses or save anything.
  3. Who sees your request. The consent sheet names where your request goes and who can read it. What will be sent? shows the exact request before it goes.

A model at home keeps your data in the house. A company's AI means your request leaves the house, so the consent sheet says so before you agree. Sometimes your admin's own server passes your request on to a company. The consent sheet then names that company: "Goes through your admin's LiteLLM proxy to Anthropic. Anthropic receives what is sent, under its own terms; your admin can read it too."

Where each feature runs

Your admin can set up two AIs at once: one at home (on the home network) and one in the cloud, and choose for each feature which one it uses, or switch it off. The app starts with the safest choice: photos, meal ideas, describing a meal, AI recipes and Ask all go to the AI at home, and a feature stays off until there is one. A feature goes to a company's AI only when your admin chooses that, and the admin screen says what it would send and where. Reading a photo of a lab report is stricter still: it runs only on the AI at home, never on a company's AI and never on a key of your own, and it is off if there is no AI at home.

Settings → AI ideas shows one line per feature, such as "Runs on your local AI" or "Runs on a cloud AI outside your network". If your admin moves a feature to a different AI, the app asks for your consent again for that AI. Your own key, if you add one, is your own choice and goes to that company. If your admin switches a feature off, it is off for everyone.

Local first, cloud only if the local answer fails. For photos, meal ideas, describing a meal and AI recipes your admin can choose this middle way. The line then reads "Runs on your local AI (host); if its answer fails the app's checks, on the cloud AI (host)". Your request goes to the AI at home first. A company's AI sees it only if the answer from home did not pass the app's checks (it was not usable, or the app's rules threw all of it out) and you agreed to send that kind of data to that company. It is not asked when the AI at home is switched off, slow, over its limit or declines the question. The consent sheet names both and has its own tick box for the company, so you can leave it off (you can agree later in Settings → AI ideas). When the company's AI answered, the answer says so: "The local AI's answer did not pass the app's checks, so the cloud AI (host) answered." Both calls show in your AI activity and count in your daily limit. Ask and lab report photos never work this way.

What AI can and cannot do here

AI can AI cannot
Pick and order meal ideas, using only foods the rules already allowed Invent a food or a nutrient number: every number comes from the app's food list
Split "describe a meal" text into search words; you pick each match Save anything on its own, unless you chose to trust it in Ask (below): you tap Add to plan or Save and the normal rules run again
Copy printed numbers from a label photo into a draft food you check, marked "from photo" Estimate nutrients from a plate photo: it gives food names and a rough weight; the numbers come from your food list
Name foods it sees on a plate photo Give doses, answer medical questions, or tell you what your lab results mean
Copy the printed results from a photo of a lab report into rows you check, on the AI at home only Say whether a result is normal, high or low, save a result on its own, or read a report on a company's AI
In Ask, answer a short question after looking up your own numbers in the app Make up a number: the app checks every number in an answer before you see it
Put together a recipe from foods the rules allowed (Recipes from AI) Set the numbers or the title: the app works them out and fits the recipe to your targets

Safety rules that always apply

  • No insulin, no doses. AI never gives insulin doses, ratios, correction factors, pump or CGM settings, or medicine advice. Carbohydrate is shown only in grams (design note 04).
  • Lows never go to AI. If you type words like "low", "hypo" or "shaky", the app shows its own "Treating a low" card and does not call AI (ADA 2026, section 6).
  • Red-flag symptoms never go to AI. If you type "chest pain", "can't breathe", "confused", "faint" or "seizure", the app shows a Get help now card (call your emergency number; treat a low first) instead.
  • Dose and lab questions never go to AI. If what you type to describe a meal also asks about insulin, a medicine or supplement dose, a lab result or what a symptom means, the app shows Ask your care team and does not call AI. Describe only the food to log the meal.
  • AI never overrides a warning. An idea that would push your day over a limit is never shown as it is. If the only problem is that a portion is too big, the app makes it smaller, a quarter of a serving at a time, until the meal fits. You then see "The app made a portion smaller so the meal fits your targets." under the idea, and the amounts and numbers shown are the smaller ones. Any other problem, such as a food to avoid with kidney disease, drops the idea. AI recipes work the same way: too much of one food in a serving is made smaller, with "The app made a portion smaller so the recipe fits your targets.", or swapped for another food, with "The app changed a food so the recipe fits your targets."
  • No free text from AI in the meal features. The reasons you read under an idea ("Low in potassium; fits your dinner carbohydrate goal") are written by the app's maintainers, not by the AI. Ask is the one place where the AI writes sentences, and the app checks them before you see them (see "Ask" below).
  • Every AI card says "AI idea · provider · model · checked against your targets · not medical advice".

Call 911 (or your local emergency number)

For chest pain, trouble breathing, a low you cannot treat by mouth, confusion or a seizure, call for help. Do not ask the app. See Get help now.

What is sent, and what is not

Sent to the AI provider Never sent
Your CKD stage, dialysis type and diabetes type ("not chosen yet" if you have not picked them; with Diabetes: None the AI is told you have kidney disease only) Your name, user name or email
Your targets and today's totals and status Your weight, height, exact age or dates
Up to 40 candidate foods with their numbers (48 for AI recipes) Entry notes, other people's data
For AI recipes: the foods and amounts of a recipe you rework, and where one serving does not fit The recipe's name, note, steps or link
Your food preferences, if you wrote any (up to 200 characters) Anything when you have not opted in
The foods and food groups you marked More like this, Less often or Not for me, by name, to meal ideas and Ask only What the app learned from your log; anything for describe a meal, photos or lab report photos
Age in 10-year bands and sex, only if you tick "share age and sex"
A label or plate photo (photo features only), with location and camera data removed
A photo of a lab report, only to the AI at home and only after you agree to it separately. The report may show your name, date of birth and record numbers, which the app cannot remove from a photo: crop it to the results table if you can
Your messages in an Ask conversation, and the numbers the assistant looks up for it (your targets, today's totals, food names and nutrients), only after you agree to it separately

Before anything is sent the first time, a consent sheet shows the provider, the server address it goes to, a line about how that provider handles data, and the exact text that will be sent. Every AI button also has What will be sent?, which shows the same thing without sending (design note 04).

Who the providers are

Your admin chooses which providers are offered. Typical ones:

Provider Where your data goes Data line shown in the app
Ollama or another model your admin runs at home your admin's computer (Ollama) "runs on your admin's hardware; your admin can read what is sent"
Hermes Agent (a dedicated, tool-free profile) your admin's Hermes, which forwards to the model it is set up with (Hermes) "your admin's Hermes agent keeps a transcript and forwards the request"
OpenAI OpenAI's servers "not used for training; abuse logs up to 30 days" (OpenAI)
OpenRouter OpenRouter, then the model's host "routed only to providers that do not collect data" (OpenRouter)

A model your admin runs at home keeps your data in the house and is the most private choice. If you have your own OpenAI, OpenRouter or Nous Portal key, you may be able to add it in Settings → AI ideas → My own AI provider (and, if your admin allows it, the address of another AI server that uses HTTPS). Your key is encrypted on the server and never shown again; changing it asks for your password. If your own provider fails, the app never sends your data to the shared one instead (Settings and keys).

Where to find it

When AI is on for you, the Add screen shows AI meal ideas, Describe a meal (AI), Read a label (AI) and, if your admin allows it, Plate photo (AI), with the number of shared AI calls left today. Each AI idea shows its foods with their warnings, the totals, what is left of your day after it, and Add to plan. If no AI idea fits your targets, the app shows its own ideas instead. Other ideas asks again without the foods it has already suggested, so you get different ones. Each ask counts as one AI call. If your admin has switched on Ask, a tab called Ask is there too.

If you change the meal, date or recipe inputs while waiting, the old answer will not appear. Ask again for your new choice. The request you already sent may still count as an AI call.

Read a lab report from a photo or a PDF

On the Lab results screen, below "Import from a spreadsheet", Read a lab report photo or PDF copies the printed results from a photo or a screenshot of your report, or from the PDF your lab's website gives you. It appears only when your admin has an AI at home that can read photos and you have turned on AI ideas; otherwise the card says why, for example "Needs a local AI on your own network, which is not set up on this server, so it is off."

  1. Take a photo, or choose a screenshot, and crop it to the results table if you can. A report can show your name, your date of birth and your record numbers, and the app cannot take them out of a picture. That is why this runs only on the AI at home.
  2. The first time, a consent sheet says where the photo goes, what the report may show, and that the photo is not kept. What will be sent? shows the request before it goes.
  3. The AI only copies what is printed. The app keeps the tests it knows (potassium, phosphate, albumin, bicarbonate, urine albumin-to-creatinine ratio, creatinine, cystatin C, eGFR and HbA1c), in units it knows, with plausible numbers, and tells you how many lines it left out.
  4. You see the rows in the same table as a spreadsheet import, all unticked, with "Read by AI from your photo. Check each value against the report before you tick it." Compare each value and its unit with the report, tick the ones that match, type the date for any row that has none, and press Save. Only ticked rows are saved. A potassium of 6.0 mmol/L or more shows the same urgent warning as when you type it.

The AI never says whether a result is normal, high or low, and it never changes your targets: only you save results, and your saved targets stay as they are until you save new ones in Profile. Talk to your care team about what a result means.

A PDF works too

Most lab websites give you the report as a PDF. Choose that file in the same card instead of a photo.

  • The PDF itself is not sent. Your browser opens it on your device and lists its pages ("Page 1", "Page 2", ...). Pages 1 to 6 start ticked; you can tick up to 10. "Reading a page takes a while on your own AI; choose the pages with results." Only the pages you tick are drawn, on your device, and sent as pictures, one at a time, to the AI at home and nowhere else.
  • Page by page. You see "Reading page 2 of 4..." and a Cancel button. The rows from all the pages come in one table, all unticked, each saying which page it came from; a row that appears twice (a header on every page, say) is shown once. If one page cannot be read, it says so and the other pages still count.
  • Some PDFs are refused, with the reason: a file over 20 MB, a file that is not a PDF, one with more than 50 pages, a damaged one, a browser too old to open it, and a PDF with a password (the app never asks for it: open it on your device and take a screenshot of each page instead). A scanned report made of pictures in a few rare formats can come out blank: use a screenshot.
  • Everything else is the same as for a photo: the consent sheet, What will be sent?, the unticked rows, the dates you type, and Save for the rows you tick.

Turn it on

  1. Check that your admin has switched AI on (Settings → AI ideas says whether it is on).
  2. Tick Use AI ideas.
  3. Choose the provider your admin offers, or your own key if allowed.
  4. Decide whether to share age band and sex. It is off by default.
  5. Optional: write food preferences such as "vegetarian, no fish" in Settings → Meal guidance → Your food preferences (the field is there when your server offers AI).
  6. The first time you use an AI button, read the consent sheet and the exact text, then accept. Photos have their own consent, and so do Ask and lab report photos. You can also ask to see the request before every AI call.

Take it back

  • Untick Use AI ideas to stop all AI calls for your account.
  • Settings → AI ideas → AI activity lists what was sent and what came back. What came back is the AI server's own answer, before the app checked it: the app used only the parts that passed, so words it never shows you, such as anything about insulin, doses or medicines, read [hidden] there. It is not advice. Your data export keeps the full text. Delete my AI history removes it. What you agreed to send has Withdraw for each server. Stored requests are kept for 30 days by default; admins see only counts, not contents.
  • If your admin changes a provider's address, everyone is asked for consent again.

Limits to keep in mind

  • AI can be wrong or slow. A home model on a small computer can take a minute or two; the app then shows its own ideas.
  • Photo reading copies what is printed. Check every field marked "from photo" against the label, and every lab report row against the report before you tick it.
  • Portions guessed from a plate photo were off by about a third on average in a 2025 test, and too small for big plates (Fridolfsson 2025). Do not count carbohydrate for insulin from a photo guess.
  • There is a daily limit on shared AI use (30 calls per person by default). One question in Ask can use several of those calls, so the limit runs out sooner.

Ask: a chat inside the app (optional)

If your admin switches on Ask, a tab called Ask appears for people who use AI ideas. You type a question, for example "How much potassium do I have left today?" or "Can I have a banana as a snack?", and press Enter.

  • The assistant looks up your own numbers in the app (your targets, your day, a food) and answers in a few plain sentences. Under the answer, From the app shows the figures the app itself worked out.
  • If your message sounds like a low, the app shows its own Treating a low card at once. If it sounds like an emergency, it shows Get help now. The AI is not asked, and nothing in Ask ever tells you to wait or skip treating a low.
  • The app checks every answer before you see it. If an answer touched on insulin or medicine doses, or had a number the app did not work out, you see a short notice instead, with the app's own numbers. Ask your care team about doses and about lab results.
  • Saving from Ask. If you ask it to log, plan or save something ("log a banana for breakfast"), it does not save it by itself. By default it shows a card: the foods the app found, the amounts and the app's totals, with the words "Nothing has been saved yet". Press Save as eaten, Save as planned or Cancel; only your press saves. You can also use the Add screen as always.
  • How much do you trust the assistant? In Settings → AI ideas you can choose Ask me before every save (the default, and the one we recommend), Trust it for this conversation or Trust it always. The last two save at once without asking, so you do not have to confirm every time, and they come with a warning first: the assistant can pick the wrong food or the wrong amount. Every number still comes from the app, the low and emergency rules and the dose checks stay on, and you can Undo each save (or delete it on Today). Photo and label readings are always shown to you to check, whatever you choose. These two options only work if your admin has allowed them; if not, Ask always asks first. "This conversation" ends when you start a new conversation or after 12 hours.
  • Telling Ask what you like. Say "I don't really like salmon" and Ask may offer to remember it for your meal suggestions. This is always a card, even if you chose to trust Ask: "Remember this for your suggestions? Salmon: less often." with Remember and Cancel. Nothing changes until you press Remember, and Undo puts back what was there. It only changes the order of suggestions, never a limit or a warning, and it is a taste, not an allergy filter: the app cannot tell whether a food contains an allergen, so check labels. Ask is told which foods and groups you marked (not what the app learned from your log), and only offers this when you say a taste in the conversation.
  • The conversation stays on the page. New conversation, reloading or signing out clears it, and the server keeps no copy of what you wrote. Your AI activity shows that a call happened, not what was said.
  • The first time, a sheet asks you to agree to send your messages and the numbers it looks up. What will be sent? shows the exact request first.
  • Ask answers can be wrong or slow, like any AI. If a number matters, check it on Today. It does not replace your care team. If a question gets no answer, it goes back into the box, and anything you typed while you waited stays there after it.

Talking to Ask (optional)

If your admin has set up speech to text, a microphone button sits next to Send.

  • Tap it, speak, and tap it again to stop (it stops by itself after a minute). Discard throws the recording away.
  • The first time, a sheet tells you where your recording goes: to the speech-to-text server your admin chose, to turn it into text, and then it is deleted. Nothing is kept. Nothing is recorded before you agree.
  • Your words appear in the question box. Read them, fix anything that is wrong, then press Send. Nothing goes to the AI until you do.
  • If your browser cannot record, or you said no to the microphone, a line tells you and you can type instead.
  • Read answers aloud (a switch under the box, off until you turn it on) speaks the answer with a voice built into your device. If your device has none, the switch is hidden: the app does not use an online voice, because that would send the answer to someone else.
  • The app does not use your browser's own speech recognition (some browsers send that audio to Google or Apple).

Talk to an assistant (optional)

Some people like to ask about food in a chat, by typing or by talking. If your admin turns on assistant access, you can connect a chat assistant you already use (for example Open WebUI, or a Hermes Agent profile) to the app with a personal token from Settings → Assistant access (Settings and keys).

  • The assistant asks the app for your targets, what is left of today, the values of a food, and how a meal would change your day. The app does the sums from your own log; the assistant is told never to make numbers up.
  • A token reads only unless you tick Can also save food and recipes when you make it. With that box ticked the assistant can also add food to your log and save a recipe as a saved meal. It is told to read back what it will save and wait for your yes, and to say what it saved. A model can still get an amount wrong, so look at Today or Saved meals afterwards; you can edit or delete anything it saved, and each save is listed in your activity. It can never change a setting or delete anything. A token that only reads leaves logging to you, in the app.
  • A second box, Also let it remember tastes (likes and dislikes), is separate from the first and also off by default. With it ticked, the assistant can remember that you like a food or a kind of food, want it less often or do not want it, and it can forget one of these. That is all it can do with it. A taste only changes what the app suggests (Not for me keeps a food out of the suggestions, though it stays in search): it never changes a limit, a warning or a number, and it is not an allergy filter (the app cannot tell whether a food contains an allergen, so check labels). The assistant is told to ask you first and never to guess a taste from what you log, but the app cannot check that. So each change is listed in your activity, a food it sets shows in Settings → Meal guidance → Your food preferences as "from your assistant" with Remove next to it (a kind of food shows in your activity, and you change it under Kinds of food in the same place), and one token can make at most 30 changes a day. Treating a low is never affected. The assistant cannot touch the Treating a low card or its options, and it cannot mark a food that treats a low, or a kind of food that has one (such as drinks or sweets), as Not for me: only you can, in the app.
  • When a message sounds like a low, the app hands the assistant the same Treating a low card you see in the app, to show first. When a message sounds like an emergency, it hands over Get help now.
  • The app cannot check what the assistant finally says. If a number looks wrong, check it on Today.
  • What you type or say, and the app's answers, go to the chat assistant and to the AI model it uses. A model run by a company sees them.

How to set it up is in the assistant guide for whoever runs your server.

If something goes wrong

  • "The AI ideas did not fit your targets today, so these are the app's own." Every AI idea failed the rules, even after the app made too-big portions smaller. Use the app's ideas.
  • "This is outside what the app's AI helps with. Your care team can answer it." The question is medical. Ask your team.
  • "This answer was not shown because it touched on medicine or insulin doses." (in Ask) The app does not show answers about doses. Ask your care team.
  • "The assistant's answer had a number the app did not work out, so it is not shown." (in Ask) The app shows its own numbers instead. Ask again, or look on Today.
  • "The assistant could not finish; here are the app's numbers." (in Ask) The AI needed too many steps. Try a shorter question.
  • AI is missing. Your admin has not switched it on, or today's shared limit is used up.

Sources

Not yet reviewed by a clinician. Facts and sources last checked 2026-10-21.