AI in your app

Your app can use AI for its own users: suggest a class, draft a reply to a ticket, summarise a week, give a student a hint. One call from a screen or a workflow returns text. The model key never reaches the browser, and you decide how much each app and each user may spend.

When to use it

• Suggestions behind a button: a class to take, a reply to send.
• Summaries of a record, a thread or a period.
• Triage in a workflow: tag or prioritise new rows before anyone looks.
• Hints with a system prompt that forbids giving the answer.

Step by step

1. Make sure AI is on for the project (Project settings, Appbricx AI). It uses your plan's credits or your own key.
2. In a component, call app.ai.ask(prompt, { system, maxTokens }) behind a user action.
3. Catch the error and show a fallback, so the screen works without AI.
4. For server-side AI, call ctx.ai.ask in a workflow or server function.

Code

The booking app's "Not sure what to take?" box:

async function suggest() {
  if (!app || !feeling.trim()) return;
  setThinking(true);
  const list = upcoming.slice(0, 20)
    .map((s) => `- ${s.title} (${s.style}, ${s.level}, ${s.duration_min} min) on ${day(s)} at ${time(s)}`)
    .join("\n");
  try {
    const answer = await app.ai.ask(`How I feel today: ${feeling}\n\nClasses this week:\n${list}`, {
      system: "You are the front desk at a small yoga and pilates studio. Recommend ONE class from the list and say why in two short sentences. Use markdown bold for the class name. Never invent classes.",
      maxTokens: 600,
    });
    setAdvice(answer.trim() || FALLBACK);
  } catch {
    setAdvice(FALLBACK);
  } finally {
    setThinking(false);
  }
}

A conversation

const next = await app.ai.chat(
  [...history, { role: "user", content: question }],
  { system: "You are a maths tutor. One hint at a time. Never give the final answer.", maxTokens: 400 },
);

In a workflow

let priority = "normal";
try {
  const a = await ctx.ai.ask(`Ticket: ${String(row.body)}`, { system: "Reply with one word: urgent, normal or low." });
  if (["urgent", "normal", "low"].includes(a.trim())) priority = a.trim();
} catch { ctx.log("AI unavailable, default priority"); }
await ctx.db.update("ticket", String(row.id), { priority });

Stay in control of cost

• On or off per project. Off means every call rejects at once with no usage.
• Model allow-list per project.
• Token budgets for the whole app and per user, over a window you choose.
• Rate limits per project, so a loop is cut off quickly.
• Usage in Project settings, Appbricx AI.

Which key pays

Add your own model key (Anthropic, OpenAI, Azure OpenAI and many more) under AI settingsand pin it to the project; your provider bills you directly. With no key pinned, calls use your plan's AI credits (see pricing). The same settings choose the key the builder's own chat uses; the full order is in AI key resolution.

Gotchas

• Always have a fallback. Calls reject when AI is off, a budget is reached or there are no credits. Never make a core action depend on an answer.
• Leave room in maxTokens. With a cap of 200 we once got an empty answer back; 600 was reliable. Treat an empty string like a failure.
• Call it on a click, not on every render. Each call is metered.
• Give it the facts. Pass the rows it should choose from and tell it not to invent others.

Related

• Custom screens
• Workflows
• AI key resolution