AI features in your App

AI your App can offer

A published App can put AI directly in front of the people who use it. Open your App in the editor, choose the Equip tab, then AI. The page has four parts, and you can use them together:

• Assistant — a general assistant your visitors can talk to, with a persona, a budget and limits you set. It is billed to your credits.
• Knowledge — reference documents the assistant answers from, so a support bot answers from your handbook and a product helper from your specs.
• Chat widget — the same assistant as a chat bubble on other websites you name.
• On-device — AI tasks that run in your visitor's own browser, with no per-use cost to you. It is always On.

Each part shows one status, and anything more to set opens in a window from its Settings icon.

The assistant, its knowledge and the chat widget belong to your published App. Until you publish for the first time their lists are empty, and pressing a switch or an add button opens Publish first, which takes you to Publish on the Manage tab. An App with nothing that can go live yet — no page, backend or action — shows one line saying so instead.

About on the Develop tab briefs the AI that builds your App. It does not change the assistant your visitors talk to; that is set here and nowhere else.

Set up the assistant

In Equip › AI, the Assistant part has a switch to turn it on, a Code icon, and a Settings icon that opens its window. As the owner, you also see what it has spent today against its daily budget. Turning it on and its budgets spend your balance, so they are yours alone: a collaborator sees them as values and can change what it says, its model and its limits. If your App's code calls the assistant while it is off, the AI tile reads Needed and the Equip tab carries a mark, so a visitor's AI feature is never silently dead.

A reel of ready-made personas runs across the top of the window — Support, Product helper, Concierge and Tutor. Pick one and it fills in the persona, which you can then edit; the lit card is the one your persona still matches word for word. Below it, the window has these sections:

• Behaviour — On, the Persona (how it behaves and what it will talk about), and the Model. Default · cheapest is chosen for you; each other option is marked Cheap, Standard or Premium.
• Budget — a Daily and a Monthly amount in dollars, and an Alerts button that takes you to where you choose whether a stop reaches you by email.
• Usage — what it has spent Today and this Month, and how many Requests it answered today. Only the owner sees this.
• Limits — Max input and Max output for a single exchange, and how many requests one visitor may make Per minute and Per hour.
• Reach — Open to reads Anyone using the app, because a published App is public.

Select Save to keep your changes. The assistant cannot be turned on without a persona and both budgets above zero; the window says Persona needed or Both budgets above zero until they are set.

An assistant stays with its App: it is instructions to the AI your visitors talk to, so it can't be starred and its persona never goes to the community. A copy of your App someone saves from the community carries no persona, so its assistant stays off until its new owner writes one, and a co-op started from your App has no in-app AI. An App you post to Slop keeps answering its visitors with your assistant as it stood when you posted, and each visitor who uses it pays for what they use, signed in to their own account — never from your credits.

Press the search on the reel of personas to find one by name; Ask AI there designs one from what you type, and its persona and limits fill the window, ready for you to Save. With nothing typed, Ask AI opens Discuss in Smart Edit.

AI usage is billed to your credits, so start with conservative budgets: anyone using your App can use the assistant up to the limits you set. Your Draft and your Live App are metered separately against the same budget, so testing on the Draft never uses up what your visitors on Live need.

Replace the assistant's connection key

Every published page carries your assistant's Connection key, so visitors' browsers can reach the assistant. It is public by design, and it is never shown in the editor.

If the key is being used somewhere it should not be, replace it. In Equip › AI, open the Assistant settings and, under Connection key, select Replace. The window tells you what happens before you confirm: the old key stops working, your live pages are moved onto the new key in the same step, and a page a visitor already has open reaches the assistant again after it reloads. If one of your pages cannot be updated, nothing changes and your existing key keeps working.

Only the App's owner sees this section.

Answer from your documents

Point your assistant at reference documents and it answers your visitors' questions from them instead of from general knowledge. When a visitor asks something, it finds the relevant passages in your documents and answers from those — so the more complete and accurate the documents, the better the answers.

In Equip › AI, the Knowledge part lists your documents with a count. Select Add reference document (the + icon) and pick a file. Reads lists what it accepts: PDF, Word, PowerPoint, Excel, Text, Markdown and CSV.

Each document reads Processing… while it is prepared, then Ready once the assistant can answer from it. A file it cannot use says Couldn’t use, with the reason beside it — a file type it cannot read, no readable text inside, or a document that could not be prepared — so you know whether to convert it or try again.

Remove a document with the trash icon beside it. The editor asks first, because the assistant Stops answering from it. To retry a failed document, remove it and add it again.

You can also ask the building assistant in plain language — "let visitors ask questions about my handbook" — and add the documents from there.

Reference documents and the read-only MotherFluxer folder

Your App's reference documents appear in a special MotherFluxer folder among your App's files, which you see under Code on the Develop tab. You can open the folder, but you can't edit it — MotherFluxer manages what it holds.

Two things are always true of this folder:

• It never holds secrets. Keys, passwords and connection details are never stored here.
• It is never part of your public App. The files your published App serves to visitors do not include it — it is material that grounds the assistant's answers, not content shipped to the open web.

To add or remove a reference document, use Knowledge in Equip › AI, never the folder itself.

Put the assistant on another website

The Chat widget puts your App's assistant on another website as a chat bubble. It needs the assistant turned on; until then its status reads Off, and once it is on it counts the sites the widget may load on.

In Equip › AI, open the Chat widget settings:

• Sites — the widget Loads only on these. Your App's own address is always listed, marked Always, and can't be removed. Type another site's address and select Allow to add it; the trash icon removes one, and the editor asks first because the widget stops loading there. A site that isn't listed never shows the widget.
• Look — an Accent colour, a Title, a Greeting, and the Position of the bubble, Right or Left.
• Code — the tag to paste into the other site's page. It changes as you set the look, so copy it last.

Run AI on your visitor's device

Some AI tasks can run entirely in your visitor's browser, on their own device — with no server, no per-use cost to you, and no data leaving the device. That makes them private, free to run, and unaffected by heavy traffic, because each visitor's device does its own work.

In Equip › AI, the On-device part is always On: there is nothing to turn on, and it draws nothing from your credits.

Ready-made starter recipes cover the most common tasks:

• Remove background — cut the subject out of a photo.
• Upscale image — enlarge an image while keeping it sharp.
• Transcribe speech — turn spoken audio into written text.
• Translate text — convert text from one language to another.
• On-device chat — a conversational assistant that runs locally.

Each recipe checks first whether the visitor's device can handle the task. When it can, the task runs locally at no cost. When it can't — an older phone, a limited browser — the recipe falls back to your App's assistant, or to a slower in-browser path, so the feature still works for everyone. You don't write that fallback; it is part of the recipe.

To add one, install it from Our features in Develop › About, or ask the building assistant for it by name — for example, "add a background remover".

How your App's code reaches AI

Your App's code reaches the assistant through built-in helpers, so you never handle an AI key yourself:

• window.mfAI — from your App's page code in the browser.
• ctx.ai — from your App's actions on the server.

Both use the assistant you set up in Equip › AI — its persona, model choice and budgets — so the same code keeps working when you change those settings, and every call draws on the budgets you set.

The Code icon on the Assistant part opens ready-to-copy calls for each way of asking: Reply (the whole answer at once), Word by word (the answer streams in as it is written), Structured (an answer in a shape you describe), and Pick a model (one call with a different model). The window also states the three facts behind every call: it is a Public endpoint, it spends Your credits, and it Stops at budget.

On-device tasks are different: they run through the recipe's own code on the visitor's device and never touch your budget.

Let the assistant look at a photo

Your assistant can answer about a picture as well as words — what is in a photo, what a label or a handwritten note says, which foods are on a plate. Your App's page uploads the photo the visitor takes or picks, then sends it with the question: up to four pictures at a time, each under 4.5 MB. Only pictures uploaded to this App can be sent.

The quickest way to add it is the AI photo reader feature: install it from Our features in Develop › About, or ask the building assistant for it. It needs the assistant turned on in Equip › AI, and Saving turned on under Storage in Develop › Data, because that is what keeps uploaded files. If the model you chose can't see pictures, the call answers that instead of guessing, so your App can tell the visitor.

When your App's AI stops answering

Your App's AI can stop for two reasons. Either way the AI tile reads Paused, the moment is recorded as an error in Manage › Health, and — if your error alerts are on — it reaches you by email, so you find out when it happens rather than when a visitor tells you.

• You hit a budget you set. The daily or monthly budget is spent, so the AI stops answering for the rest of that day or month. Raise the budget in the Assistant settings in Equip › AI, or wait for the period to roll over. As the owner, you can see a stop coming: Manage › Alerts shows how much of each budget your Live App has used.
• The App passed a platform safety limit. An App making an extraordinary number of AI requests in one day stops answering for the rest of that day, to protect your credits from a runaway loop. It starts again on its own the next day. If whatever caused the surge is still running, it will stop again, so fix the cause first.

While the AI is stopped, anything in your App that asks it reports that it's unavailable. The rest of your App keeps working.

Ask the building assistant to add it

You don't have to wire any of this up by hand. Describe what you want to the building assistant in the App editor, in plain language, and it builds it in:

• "Add a background remover so visitors can cut out photo subjects."
• "Let visitors ask questions about my handbook."
• "Add a translator to this page."
• "Give my App a chat assistant that answers product questions."

It adds the right recipe or wires up the right helper. Refine it the same way — describe the change you want.