Intelligence (AI Capabilities)

Intelligence is Al Islam's model layer. It is not a chat window bolted onto the side of the app — it sits underneath every reading surface, so the app's own content is what the models actually work on: the ayah, the narration, a scholar's translation, a tafsir, your own notes.

It runs on your models, your accounts, your keys. Al Islam never resells model access, and your keys never leave your device.

Introduced in version 4.5 ("Bayt al-Hikmah"). Part of Al Islam Premium.

Contents

• The Idea
• Quick Start
• Part I — Connecting a Provider
• Part II — What Your Models Can Do
• Part III — The Reading Layer
• Part IV — Standing Translations
• Part V — Your Own Commands
• Part VI — Dictation
• Part VII — Routing
• Part VIII — The Screens
• Part IX — Engine Settings
• Part X — Keys, Traffic and Privacy
• Part XI — What It Costs
• Part XII — When Things Fail
• Part XIII — Design Principles
• Part XIV — Limits and Honest Caveats
• Frequently Asked Questions
• Glossary

The Idea

Most apps that add AI do one of two things. They open a chat box in a corner, leaving you to copy your content into it. Or they resell somebody else's model at a markup, taking your content through their servers on the way.

Intelligence does neither.

The work happens where you are reading. Open a hadith and the actions are already beneath it. There is nothing to copy, nothing to paste, nothing to navigate away to. The model receives the passage, its reference, and its language — because the app already knows all three.

The account is yours. You connect providers you already have, using keys you already hold. Whatever your provider charges is between you and them. There is no shared account, no token bundle, no markup, and no metering.

That second decision is the more consequential one, and it shapes everything else:

The premium subscription unlocks the suite — the routing, the reading layer, the library, the studio, the dictation. The usage runs on your terms.

Quick Start

Three steps, once.

1. Connect an account. Pick your provider from the catalogue and paste your key. It is verified on the spot, so a mistyped key is caught immediately rather than the first time you try to read something.

2. Choose what serves what. Pick a model for translation, for explanations, for speech. One choice is enough to begin — everything else falls back to it.

3. Open any hadith or ayah. The actions are already there, underneath the text.

Nothing else needs configuring. Everything beyond this point is refinement.

Part I — Connecting a Provider

The catalogue — 35 services, 10 protocols

Al Islam speaks ten different provider protocols, covering thirty-five services between them.

That last row matters more than its size suggests. The general connection means a service that did not exist when this release shipped can still be used today, as can a company proxy, a self-hosted gateway, or a regional provider we have never heard of.

What a connection asks for

Every service has its own idea of what a connection looks like — a key; a key and a region; a key and an endpoint; a project id; a deployment name; nothing at all.

Al Islam builds the right form for whichever service you choose, so each asks for exactly what it needs and nothing more. You are never presented with an endpoint field for a service that has a fixed endpoint, or a region selector for a service that has no regions.

Instant verification

Saving a connection tests it straight away, against the lightest call the service offers. A mistyped key, a wrong region, an endpoint with a typo — all caught at the moment you paste, not the first time you try to read something.

As many accounts as you like

A personal key and a work key. A fast inexpensive model and a careful expensive one. A cloud service and a local one. Each connection is enabled or disabled on its own, so you can park one without deleting it.

Custom headers

Every connection accepts custom headers, passed through exactly as written — enough to reach a company proxy, an authenticating gateway, or a service with its own conventions.

Local models

Five of the supported options — Ollama, LM Studio, Jan, llama.cpp and vLLM — run on your own machine. No key, no account, no bill, and no data leaving your network at all.

For anyone who wants the capability with none of the exposure, this is the configuration to choose. It is also the only configuration in which Intelligence works fully offline.

Masked once saved

A key is shown only as a hint (••••3kfa) and never displayed again. See Part X for the complete account of how credentials are handled.

Part II — What Your Models Can Do

Four capabilities, each routed independently.

Not every service does all four, and the app knows which does what. The catalogue shows each service's capabilities before you connect it, and the routing table only offers a connection for work it can actually perform.

Part III — The Reading Layer

The screens are the smaller half of Intelligence. The larger half appears underneath the content itself.

Every reading surface in the app — a hadith, an ayah, a scholar's translation, a tafsir, a translation your own model produced, and the random hadith on the home screen — carries the same row of three actions.

Translate

One tap renders the passage into the language you choose.

• It looks like a translation. The result is presented as a scholar-style card, laid out exactly like a published translation — the same typography, the same per-language fonts you have already set in Text Customization, the same reading experience. A translation reads as a translation regardless of what produced it.
• Translations are kept. Every one is saved to your Library, attached to the ayah or narration it came from, so you never pay twice for the same passage and never lose one you wanted back.
• Duplicates are recognised. Asking again for the same passage with the same model returns what you already have.
• Repeats are instant and free, served from the device rather than the network.
• Changed sources are noticed. If the underlying text is later updated, the saved translation is marked out of date, with one tap to produce a fresh one.
• Pin the ones that matter so they stay at the top of your Library.
• Copy or insert. Take the text out, or drop it straight into a note.

Translation lanes

A lane turns a good combination into a single tap. Save one as Urdu via DeepL or English via Gemini — a target language, and optionally a specific account and model — then choose it from the reading itself, without ever opening the routing table.

Your last choice is remembered, so the common case is one tap from then on.

Listen

Content is read aloud in a natural voice, and listening reaches everything in the app with words in it — not only the app's own text.

• Readings — the Arabic, or any translation shown beneath it.
• Answers — anything a model has written for you, spoken once it is finished rather than half-written.
• Your notes — a Qalam document is read as continuous prose, so headings, formatting marks, table borders and drawings never reach the voice.
• Voice and speed are yours, adjustable from the player and remembered between visits.
• Speech can have its own provider. A specific account and model can serve speech regardless of what serves everything else — useful when one provider has the voice you like and another has the translations you trust.
• One voice at a time, and it keeps reading when you leave the screen, with the player following you.

Ask

A conversation about the passage in front of you, appearing inside the screen you are already reading — nothing to copy out, nothing to navigate away to.

• Threads stream in as they are written.
• Follow-up questions keep their memory, so a conversation builds instead of restarting.
• Answers can be read aloud, copied, or inserted into a note.
• Each reading keeps its own thread, so moving between ayat does not mix them up.

The instructions behind Ask are written specifically for Islamic source material: stay faithful to the text, invent no rulings and no attributions, preserve honorifics, and say plainly when something is uncertain rather than smoothing it over.

This reduces a class of failure. It does not eliminate it — see Part XIV.

The workspace

For longer work, the same three abilities open into a fuller surface with Translate, Ask and Listen side by side as independent sections — translate a passage, question it, and hear it without any one of them interrupting another.

Where a reading has several versions on screen — the Arabic and its translations, or a narration and its renderings — each section chooses its own source. Ask about the Arabic while listening to the Urdu, or translate a scholar's English into a fourth language while the original stays in view.

Part IV — Standing Translations

Beyond one-off actions, a language can be pinned to stay.

Set a standing translation to Urdu, and every hadith you open arrives with Urdu already beneath it — filled in as you go, laid out as an ordinary part of the page rather than something you had to ask for.

• More than one at a time. Urdu and English together, or three languages one below another.
• Chainable. Take a scholar's English rendering and produce Turkish from it, or build a second language out of the first rather than out of the original.
• They report themselves. If a connection is removed or a model stops answering, the card says so plainly instead of quietly disappearing.
• They know where they belong. A standing translation built from one book's published translation appears only in that book. One built from the narration itself works everywhere — including the home screen, where the hadith belongs to no particular book at all.
• Reorder, rename and remove them at any time from the reading itself.
• They work in the Quran too, filling in as you move from ayah to ayah.

Part V — Your Own Commands

Translation and asking cover most of it. For everything else, you can write your own instructions and run them over any passage in the app.

A command is a short template with placeholders for the passage, its reference, and a target language:

Extract the fiqh points as a numbered list. Explain this to a ten-year-old. Give me three lessons and one action for today. Rewrite this translation in plainer English.

Write it once and it appears wherever readings appear.

Five are included to begin with — Summarize, Key lessons, Explain simply, Deep explanation and Remaster — and every one of them is yours to rewrite or delete.

Commands and translation lanes are both managed in the Actions studio, where you can see what each will do before you run it.

Part VI — Dictation

The Qalam notes editor gains a microphone.

Tapping it opens a recorder inside the editor — nothing opens over your note, nothing navigates away from it.

• Record, pause, resume, and stop when you say so. It never decides on its own that you have finished speaking. There is no wake-word, no background listening, and no silence detection that starts or stops capture for you.
• A live waveform driven by your actual microphone level, so you can see that you are being heard.
• A running timer and an unmistakable state at every stage.
• Import instead of recording — hand it an audio or video file you already have and transcribe that.
• Transcribe is a deliberate step. When you take it, the words land exactly at your cursor, inside the note you were already writing.
• Language can be stated or left open for the model to work out.
• Record as many times as you like into the same note, each take landing where you put it.

Dictate a lecture as you listen to it, capture a thought on the move, or write a long note without typing a word of it.

The editor also gains Ask Intelligence on a selection or on the whole note — question what you have written, or have a passage explained, with the answer available to insert straight back into the page.

Devices with no microphone can still transcribe an imported file.

Part VII — Routing

The Routing screen decides which connection, model and voice serves each kind of work — and lets you override it for individual areas of the app.

• Read hadith with one model, the Quran with another, your notes with a third.
• Anything you have not set falls back to a single global default, so no request is ever left unanswered.
• Model and voice lists are read from your own account wherever the service supports it, so you see exactly what you are entitled to use — not a hardcoded list that goes stale.
• Any model id can be typed in directly. A model released this morning is usable this morning, without waiting for an app update.
• Per-capability tuning — creativity, response length, formality, and the other settings each service genuinely accepts — shown where it applies and hidden where it does not.

The design intent is that routing is something you can use, not something you must. A single global default is a complete configuration. The table exists for the person who wants a cheap fast model summarising and a careful expensive one translating.

Part VIII — The Screens

The Playground deserves a note. It exists so you can evaluate a model without spending your own content on the experiment — try a voice before assigning it, compare two models on the same phrase, confirm a new connection works end to end. Nothing you do there touches your readings or your notes.

The Activity screen exists so the bill is never a surprise. It shows what ran, on which connection, how long it took, and how much was used — before your provider's invoice tells you the same thing a month later.

Part IX — Engine Settings

Every one of these has a sensible default. The ten-minute ceiling exists for slow local models on modest hardware; the five-second floor exists for anyone who would rather fail fast than wait.

Part X — Keys, Traffic and Privacy

A system holding your provider credentials owes you a plain account of how it treats them.

Your keys

• Keys are encrypted on your device with a hardware-backed key that cannot be taken off it, held separately from everything else the app stores.
• They are never shown again after you save them — only a masked hint.
• They never appear in logs, usage records, backups or exports. This holds on every plan, under every setting, with no exception.
• Removing a connection removes its keys first, so a connection can never be left behind without a way to clear it — and removal stays available to you at all times.

Your traffic

• Traffic goes only where you sent it. Al Islam's AI connection has no default destination; requests reach the providers you configured and nowhere else.
• We operate no proxy and no relay. Your device contacts your provider directly. We cannot read your prompts, your content, or the responses — not because we promise not to, but because they never travel anywhere we can see them.
• Nothing is sent anywhere until you ask for it, and when you do, it goes to the provider you chose.

Your records

• Usage records hold facts, not content — which service, which model, how long it took, how much was used.
• An excerpt of what you read is stored only if you switch that on yourself. It is off by default.
• Caching and usage recording can both be switched off entirely.

Provider messages

Some services quote your own credential back inside their error text. Anything of that shape is stripped before it reaches the screen, so a provider's carelessness cannot become your exposure.

What this does not cover

Once your request reaches the provider you chose, that provider's privacy policy governs it, not ours. Providers differ significantly in whether they retain submitted data or use it to train models. We strongly encourage reading the policy of any provider before connecting it — and the catalogue links to each service's own documentation for exactly this reason.

If you want the capability with none of that exposure, use a local model. Nothing leaves your machine.

The complete account of the app's data handling is in the Privacy Policy.

Part XI — What It Costs

Al Islam Premium unlocks the Intelligence suite. Your provider bills you for the models.

We do not resell model access, mark it up, meter it, or take a share. There is no token bundle to buy from us and none to run out of.

What that means in practice:

• A local model costs nothing at all. No key, no account, no bill.
• Translation services are typically the cheapest paid route for translation specifically, and several have free tiers.
• Inference platforms often serve open models at a fraction of major-lab pricing.
• Major labs cost the most and are usually the most capable.
• Repeat translations cost nothing, because they are served from your device rather than the network.

The Activity screen shows what you have used so you are never guessing, and every service in the catalogue links to its own console so you can check pricing before you commit.

Part XII — When Things Fail

Model providers fail in ordinary ways — a rate limit, an expired key, a model withdrawn, a network that dropped. Intelligence is built to say which.

• Failures are named, not swallowed. Each provider's own error type is mapped onto a common vocabulary, so a rate limit reads as a rate limit rather than as a short or empty answer.
• Retries are bounded and configurable, and only failures that could plausibly succeed on a second attempt are retried.
• Streaming answers are held back until the text stops arriving where partial output would mislead, so you never read half a sentence as though it were the whole one.
• Standing translations report their own breakage. If the connection behind one is removed, the card says so rather than quietly vanishing.
• Connection health is visible on the Hub and on each connection's detail screen, so a key that stopped working is visible before you depend on it.

Part XIII — Design Principles

Five decisions shaped the system, and each has a reason worth stating.

1. Underneath the content, not beside it. A chat window makes you the integration layer — copying passages out, pasting results back, losing the reference on the way. Putting the actions beneath the reading means the app supplies the passage, its reference, and its language automatically. The model gets better context than you would have pasted, and you do no work.

2. Bring your own key, without exception. Reselling model access would have been simpler to build and easier to monetise. It would also have meant your content passing through our servers, your costs hidden inside our margin, and your model choice limited to what we had negotiated. Every one of those is worse for you.

3. Presets never cage. Wherever the app offers a list — models, languages, voices, commands — it also accepts anything you type. A model released this morning works this morning. A language not in the list is still a language. The list is a convenience, never a boundary.

4. Output must look like the app. A translation is presented as a scholar-style card in your own per-language typography, not as chat output in a bubble. Content produced by your model is content, and it should read like the published translations sitting beside it.

5. Say what happened. Failures are named. Stale translations are marked. Broken standing translations announce themselves. Usage is visible before the invoice arrives. A system that spends your money and speaks about your religion should never be vague about what it just did.

Part XIV — Limits and Honest Caveats

This section matters more than any other in this document.

Models fabricate, and they do it fluently

Language models produce confident, well-formed, entirely invented output. In Islamic material this shows up as invented rulings, invented attributions, and citations that look authentic but are not — a fabricated chain of narration reads exactly like a real one to anyone who cannot check it.

This is a property of the technology. No prompt, no provider, and no app removes it.

What Al Islam does about it

• The instructions behind Ask are written for Islamic source material specifically: stay faithful to the text, invent no rulings and no attributions, preserve honorifics, and state uncertainty plainly rather than smoothing it over.
• Model output is always presented alongside the authentic sources it came from, never in place of them.
• The app carries 232,559 hadith and 122 scholarly translations offline, so the authentic reference is always one tap away.
• Translations are attributed to the model that produced them, and saved with that attribution.

These measures reduce the failure rate. They do not make output trustworthy on its own.

What you must do about it

Verify anything consequential against the authentic sources in the app, or with a qualified scholar.

AI output is not a religious authority. It is not a fatwa, not a scholarly verdict, and not a substitute for qualified Islamic scholarship. Treat it as a study aid — a fast first pass, a rough rendering, a prompt for your own thinking — and never as a final word.

Other limits worth knowing

• Translation quality varies enormously between models, languages, and passages. Classical Arabic is harder than modern prose, and a cheap model will show it.
• Speech mispronounces Arabic to varying degrees depending on the voice and provider.
• Transcription accuracy depends on audio quality, accent, and background noise.
• Provider availability, pricing and model behaviour are outside our control and may change without notice.
• Intelligence requires a connection, unless you are running a model locally.

Frequently Asked Questions

Do I need Intelligence to use Al Islam?

No. Every core feature of the app — Quran, Hadith, Prayer Times, Qibla, Hijri Calendar, Tasbeeh, Notes, Search, Collections, History — works exactly as it always has, with no model connected and no subscription. Intelligence adds a layer; it replaces nothing.

Why do I need my own API key?

So your content never passes through our servers, your costs are never hidden inside our margin, and your choice of model is never limited to what we negotiated. See The Idea.

Can I use it for free?

Yes, in two senses. Local models (Ollama, LM Studio, Jan, llama.cpp, vLLM) cost nothing to run and need no key. And several hosted services have free tiers sufficient for occasional use. A premium subscription is still required to unlock the suite itself.

Which provider should I choose?

For translation specifically, a dedicated translation service is usually cheapest and often best. For explanation and discussion, a major lab model. For Arabic and Islamic material, Fanar is built for it. For privacy above all, a local model. For cost, an inference platform serving open models.

Nothing stops you connecting several and routing each kind of work to the one that suits it.

Is my key safe?

Keys are encrypted on your device with a hardware-backed key that cannot be taken off it, shown once and then only as a masked hint, and excluded from logs, usage records, backups and exports. Removing a connection removes its keys first. See Part X.

Does Al Islam see what I ask?

No. Requests go from your device to your provider directly. We run no proxy and no relay, so there is no point at which your prompts or responses are visible to us.

Does my provider see what I ask?

Yes — that is who you sent it to. Their privacy policy governs what happens next, and providers differ significantly in whether they retain or train on submitted data. Read theirs before connecting, or use a local model.

Can I trust an AI translation of the Quran?

Treat it as a study aid, not as a scholarly translation. The app ships 122 published translations by named translators precisely because those are the ones to rely on. An AI rendering is useful for a language no published translation covers, or for a quick sense of a passage — always with the authentic text beside it.

What happens to my translations if I cancel premium?

They stay. Saved translations, commands, lanes and notes are your content and are not deleted when a subscription ends. Premium features become unavailable; your data remains yours and remains on your device.

What happens if I remove a connection?

Its keys are removed first, then the connection. Any standing translation or route that depended on it reports itself as broken rather than silently disappearing, so you can reassign it.

Can I use a provider that isn't listed?

Yes. The general connection accepts any OpenAI-compatible gateway, proxy or service, with custom headers passed through exactly as written.

Does it work offline?

Only with a local model. Translations already produced are kept on your device and remain readable offline regardless.

Glossary

BYOK (Bring Your Own Key) — a model in which the app provides the interface and you provide the provider account. Your credentials, your billing, your choice.

Capability — one of the four kinds of work: translation, text generation, speech, transcription.

Connection — one configured provider account. You may have as many as you like.

Lane — a saved translation combination (a target language, optionally a specific account and model) reachable in one tap from a reading.

Local model — a model running on your own machine rather than a remote service. No key, no bill, nothing leaving your network.

Protocol — the API convention a service speaks. Al Islam speaks ten, which is how thirty-five services are supported.

Routing — the table deciding which connection, model and voice serves each capability in each area of the app.

Standing translation — a language pinned to appear automatically beneath every reading, rather than requested one passage at a time.

Related Documentation

• Notes — Qalam — the editor Intelligence writes into
• Text Customization — the typography translations adopt
• Hadith and Quran — the reading surfaces the layer sits beneath
• Privacy Policy — the complete account of the app's data handling
• Changelog — what shipped when

Questions? Email support.alislam@lymrah.com, or reach us on X or Instagram.