AI writing

relaybell can write the words of a notification for you. It is AI: normally a small language model that runs on relaybell's own servers, not a third-party AI service (see where it runs). There are three ways to use it:

• Write it for me (the composer): say in a sentence what the notification is about, pick a tone and, if you like, a product or article from your catalog. You get 3 versions to choose from, then edit the one you pick.
• Improve my text (the composer): 3 clearer, more clickable rewrites of the title and message you wrote.
• AI tailors the wording for each person (a send, an automation or a journey message): each person gets a version written around their interests, and a share of people get your original so you can compare. See AI-written wording.

The AI can get things wrong. Read what it wrote before you send it.

What the AI sees

Only what it needs to write the copy:

• your title, message and the sentence you typed,
• your website's name and language,
• the product or article you picked (its name, price, category and summary), and
• for wording per person, interest names such as "Running shoes" or "Camping".

It never receives who a person is: no email, name, user id, location, browser, history or anything else from a subscriber profile, and the text isn't used to train anything.

Where it runs. The composer's privacy line and GET /api/v1/ai/status (provider) say which applies to relaybell right now:

• relaybell — the model runs on relaybell's own servers. Nothing above leaves relaybell's servers.
• anthropic — relaybell is set to use Anthropic's AI (Claude). The items above are sent to Anthropic to write the copy; still nothing about your subscribers.
• external — another AI service. The items above are sent to it; still nothing about your subscribers.

What it is never allowed to write

Every version is checked before you see it, or before anyone receives it. A version that breaks a rule is thrown away; if none is left, you get a plain message, and a send goes out with your original words.

• No made-up facts. Every number, percentage, price and amount must already be in your text, your sentence or the product, written in digits or in words ("ten percent", "three days"). Sales, discounts, "% off", "save $20", coupon codes and anything "free" need the same kind of offer in your own title, message or sentence (or the product's price); the same goes for deadlines ("last chance", "ends tonight") and low-stock claims ("selling fast", "only 3 left") — "back in stock" doesn't allow "selling fast". Days, months, seasons and holidays ("summer", "Diwali") must be in what you gave. If you want "20% off this week only" in it, say so.
• No links, email addresses or phone numbers. The link is the notification's own url.
• Merge tags stay as you wrote them. The AI never sees your tags: each one is swapped for a neutral mark before it writes and put back after. Every version keeps each tag of your copy exactly as written ({name|there}, recommendation tags like {rec.title}) and as often as you used it; a version that drops, repeats or changes one, or writes a name or [there] in its place, is thrown away.
• No shouting or spam. No words in capitals you didn't use, no "!!", "act now", "click here" or "you've won", at most 2 emojis.
• Plain, clean text within 65 characters for the title and 140 for the message, in the language you wrote in — or, when that isn't clear, your website's language (the most common language of your subscribers).

Limits

• 500 AI versions per website per day (UTC). Each "Write it for me" or "Improve my text" counts as one; for wording per person, each version written counts as one. The same request again within 7 days reuses the earlier answer and isn't counted. When the day's allowance is used up, the buttons say so, and sends go out with your original words until midnight UTC.
• 30 writing requests an hour from one IP address (rate limits).
• Writing takes a few seconds (about 10–20 s for 3 versions). When the AI is busy with other requests or slow, you get a "try again in a minute" message instead of a long wait.

API

The same routes the composer uses. Session or an API key with the send scope (scopes).

GET /api/v1/ai/status — is AI writing available for this website right now?

{ "available": true, "provider": "relaybell", "dailyLimit": 500, "usedToday": 12 }

provider is relaybell (the model on relaybell's own servers), anthropic (Anthropic's AI), external (another AI service) or off — see where it runs.

POST /api/v1/ai/write — 3 versions from a short description.

curl -X POST https://relaybell.com/api/v1/ai/write \
  -H "Authorization: Bearer rb_live_xxxxx" -H "Content-Type: application/json" \
  -d '{ "brief": "Our autumn jackets are in. 20% off this week only.", "tone": "friendly" }'

{
  "options": [
    { "title": "Autumn jackets are here", "body": "20% off every jacket this week only. Find your new favourite." },
    { "title": "New jackets, 20% off", "body": "Warm, light and ready for chilly mornings. This week only." },
    { "title": "Your autumn layer is in", "body": "Take 20% off our new jackets this week." }
  ],
  "provider": "relaybell",
  "model": "qwen2.5:3b"
}

Fewer than 3 versions come back when some didn't pass the checks.

POST /api/v1/ai/suggest — "Improve my text": { "title", "body", "tone"?, "language"?, "variant"? } → { "ai": true, "suggestions": [{ "title", "body" }], "provider", "model" } with up to 3 rewrites (language left out: the language of your title and message, else your website's). When AI writing is switched off for relaybell it still answers 200 with "ai": false and 3 simple rewordings of your own text.

Errors — each with a plain message you can show as it is: