Website chat
One script tag. Under 3 KB before anyone opens it.
كليم
Kalim is a shared inbox for WhatsApp, Instagram, Messenger and your website chat, with an AI support agent. It answers your customers in Arabic or English, cites the article the answer came from, and hands the conversation to a person the moment it cannot.
No card. One snippet. Your team is answering in about ten minutes.
Your customers write where they already are — WhatsApp at midnight, Instagram after seeing a post, Messenger from your Page, the chat on your site. Your team answers all of it from the same screen, in the same thread, with the same saved replies and the same AI.
One script tag. Under 3 KB before anyone opens it.
Cloud API. Replies inside the 24-hour window cost nothing.
DMs from the account linked to your Page, answered like any other message.
Messages to your Page, with the sender's name rather than an id.
A bot token is the whole setup; Kalim registers the webhook.
An address customers write to, threaded into conversations.
Every channel is on every plan, free included. Connecting one is a page id, a token and a paste of the webhook URL — the screen that asks for them shows both.
Not a bubble that opens an empty box. It knows who is writing, keeps their thread between visits, and gives your agents everything they need to answer in one screen.
Pass your user's id, name, plan or anything else you define, signed on your server. Your agent sees it before the customer types a word.
Not three dots — the actual words, as they write them, and the deletions. You are already drafting the answer before they hit send.
Close the tab, come back next week, same conversation. A chat ends when your team resolves it, not when a customer closes a window.
Customers send the error instead of describing it; you send the annotated fix. Images and video play in the thread rather than downloading.
Arabic and English in the same conversation, each message reading in its own direction — including the box your customer types into.
2.7 KB on page load; the chat itself downloads on the first click. It rings and notifies when you reply, even if they closed it.
Paste it before </body>. The loader is 2.7 KB gzipped and the chat
itself only downloads when somebody actually clicks it — your Lighthouse score will not
notice we are there.
Tell it who is chatting and your agents see the customer's name, plan and order id before they type a word. You decide which of those fields are required; an install missing one is refused at install time, not discovered months later in a half-empty CRM.
Sign up, confirm your email, and the website chat channel is already there with an id of its own.
One script tag before </body>, carrying that id. The chat is
live on the next page load.
Your server signs the fields it sends with your workspace secret, so an agent can trust the name and order id in front of them.
<script>
window.KalimSettings = {
websiteId: 'your-website-id',
identity: {
externalId: 'user_8814',
firstName: 'Nadia',
attributes: { tier: 'Gold' },
v: 2,
hmac: '…signed on your server',
},
};
</script>
<script src="https://cdn.kalimchat.com/loader.js" async></script>
Most support AI is a black box with a credit meter attached. Ours shows the bill, the sources, and the choice of engine — including running an open-source model on your own hardware, where the only cost is electricity.
An answer that cannot point at a help-centre article is never sent. It becomes a handoff with the draft attached, because a confident wrong answer costs more than a queued ticket.
Run the same question through Claude and an open-source model side by side, read both replies, and keep the one that suits your customers. Then route Arabic to one and English to the other if that is what wins.
Cost per resolved conversation, measured — not estimated — next to what it should be and what the alternatives charge.
| Kalim, routed and cached | $0.026 | Small model first, warm prompt cache |
|---|---|---|
| Kalim, on your own hardware | $0.00 | Open-source model, your GPU, your data |
| A large model, no caching | $0.081 | Full prompt re-sent every turn |
| What Crisp charges | $0.056 | Derived from their credit tiers |
Built here, not translated here
Most tools bolt right-to-left on at the end and it shows: the caret starts on the wrong side, dates come out backwards, and the AI answers formal Arabic to someone who wrote in their own dialect.
Kalim types right-to-left per message rather than per app, so one thread holds both scripts without either looking like a mistake. The AI answers in the language, variety and register the customer used — Modern Standard or dialect — and you can point Arabic conversations at a model chosen for Arabic.
Rankwell and TrueAvis are two products we also build, and both run their support on Kalim — in French, Arabic and English, from one inbox. If it could not handle that, we would know before you did.
Our support was a shared mailbox and two people answering over each other. Kalim gave us one thread per customer and a saved reply for the audit questions we get every week, so a first answer goes out in minutes instead of whenever somebody next opens the tab.
We sell to merchants in French and Arabic, and answering both from one inbox without switching tools is the part that actually saved us time. The AI drafts the reply and cites the help article, and we send it or fix it.
Charging per seat makes a growing team think twice about adding the person who would answer faster. So the paid plans are one price however many people you put in the inbox — what changes is how much of the work you do yourself.
$0/month
One person, answering their own customers.
$19/month
Everyone in one inbox, flat.
Everything
$49/month
The parts a person should not have to do.
Every plan gets every channel. The AI stops rather than overcharging you, and we never bill for it without asking first.
Three differences worth knowing before you migrate, and the things Kalim does not do yet.
Kalim is $19 a month for the whole team — not per seat. Hiring a second agent, or adding the person who only answers on Saturdays, costs nothing. Most support tools bill per seat, which quietly prices you out of putting more people on the queue at exactly the moment you need to.
Every AI reply has to point at the help-centre article it came from. When it cannot, the conversation goes to a person with the draft attached rather than a confident guess going out under your name. That is a design decision, not a setting — there is no way to switch the citation gate off.
Each message sets its own direction, so an Arabic question and an English reply sit correctly in one thread. Arabic conversations can be routed to a model chosen for Arabic while English stays on another. The agent's own inbox runs in Arabic, French or English, mirrored properly.
Kalim speaks the OpenAI-compatible API that Ollama, vLLM and TGI serve, so an open-weights model on your own GPU is a first-class engine rather than a fallback — your conversations never leave your machines. You can bring your own Claude key instead, and compare the two on cost and citation behaviour first.
The longer comparison, including what to check before you migrate →
If yours is not here, the chat in the corner is the product — ask it, and a person picks up whatever it cannot answer.
The only question is which inbox it lands in.
Start freeThe chat in the corner is this product. Ask it something.