ضحى duha · closed beta
Duha is an AI workspace that answers questions about your deen from the Qurʾān itself, two complete tafsīrs, every ṣaḥīḥ ḥadīth in Bukhārī and Muslim, and the life of the Prophet ﷺ. Every answer cites exactly what it stands on, or tells you honestly that it can’t.
closed beta · sign-ups open in waves. leave your email and we’ll write when the window opens.
Everything you’d expect from a frontier AI workspace: agents, deep research, documents, email, your calendar. And one line it will never cross — on matters of deen, it answers from verified sources, or not at all.
36,503 verified records · answer-quality gate green on every deploy · cite-or-refuse enforced in code · no ads, no trackers
how it works
Most AI will answer any religious question with total confidence and zero sources. Duha is built the other way around: the honest path is enforced in code, not requested in a prompt.
Answers come only from a byte-verified corpus of canonical texts, never from model memory. If retrieval finds nothing, there is no fallback path.
Every answer must name real sources (collection, number, kitāb) or the pipeline rejects it and says plainly that no verified source covers it.
Personal rulings (divorce, inheritance, oaths) get a referral to a qualified scholar, not an AI verdict. Distress gets real-world help, always.
Duha’s answers stand on 36,503 verified records: every āyah of the Qurʾān with two complete English commentaries, every ḥadīth in Ṣaḥīḥ al-Bukhārī and Ṣaḥīḥ Muslim, the complete Mukhtaṣar al-Qudūrī, and the Prophet’s ﷺ life from four published biographies. Each one is checked character-by-character against canonical sources on every build. If an answer can’t be cited, Duha says so, and points you to a qualified scholar instead. والله أعلم — and Allah knows best.
Every source shows its true verification status. Today the corpus is machine-verified; the scholar review is underway — and until a named scholar signs off, every answer says so.
the library
Not the open internet, and not model memory. A closed shelf of canonical texts, ingested character-by-character, verified on every build, and growing carefully.
new · qurʾān & tafsīr
Name a sūrah, or just ask what an āyah means. Duha walks it in order: the Arabic verbatim, the translation, and what the mufassirūn actually said, with every passage cited to its commentary.
بِسْمِ ٱللَّهِ ٱلرَّحْمَٰنِ ٱلرَّحِيمِ
With the name of Allah, the All-Merciful, the Very-Merciful.
ٱلْحَمْدُ لِلَّهِ رَبِّ ٱلْعَٰلَمِينَ
Praise belongs to Allah, the Lord of all the worlds.
ٱلرَّحْمَٰنِ ٱلرَّحِيمِ
the All-Merciful, the Very Merciful.
…and on through 1:7: one answer, every verse in order, 13 citations.
closed beta, honestly labeled
we’d rather tell you what’s unverified than pretend it isn’t.
the workspace
Not a fatwa hotline — a daily driver. And the Islamic layer runs everywhere: ask a deen question mid-conversation and it gets grounded, cited, or referred, without switching apps.
Every message in every chat runs through the same safety-critical classifier — religious questions get grounding and citations, personal rulings get referred, distress gets real help. General questions flow through untouched.
Bring any OpenAI-compatible endpoint, self-hosted or private. When a model gets rate-limited, the chain falls back to a sibling — never to a fabricated answer.
Hand it real work: shell, files, web search, scheduled tasks, your own MCP servers and skills.
A writing-first editor with AI edits and suggestions, drafts, PDFs and forms.
Goals, reminders, CalDAV sync, event chips right in chat.
Your own IMAP inbox: triage, summaries, reply drafts, attachments into chat.
Multi-step web research that reads sources and renders cited reports.
Add Duha to your phone’s home screen or laptop dock straight from the browser — no app store. It opens in its own window, keeps working offline, and your data stays on infrastructure you control.
family
Child-safety is enforced on the server, not hidden in a settings menu.
why trust it
the reasons behind the design, not just the features.
Muslim-built, methods in the open: a public threat model, a security policy, and a public correction trail for the corpus itself. Known source artifacts are flagged, never silently “fixed.”
No Arabic is ever typed by a human or a model — every character is a programmatic slice of canonical digital sources, and a validator proves it on every single build.
Self-hosted, open-weight models, your data on infrastructure you control. Your religious life is not a data product. No ads, ever.
مَا وَدَّعَكَ رَبُّكَ وَمَا قَلَىٰ
“Your Lord has not forsaken you, nor does He despise you.”
Sūrat aḍ-Ḍuḥā · 93:3
questions, answered
No, and it refuses to act like one. General questions get grounded, cited answers. Personal rulings (divorce, inheritance, oaths, disputes) get what the sources say in general plus a referral to a qualified scholar — never an AI verdict. That boundary is enforced in the pipeline, not requested in a prompt.
A closed corpus of 36,503 verified records: the whole Qurʾān (Uthmani text with English), two complete tafsīrs (Ibn Kathīr and Maʿārif al-Qurʾān), every ḥadīth in Ṣaḥīḥ al-Bukhārī and Ṣaḥīḥ Muslim, the complete Mukhtaṣar al-Qudūrī, and four published seerah biographies. Never the open internet, never model memory. If retrieval finds nothing, Duha says so.
Fiqh answers currently stand on Ḥanafī texts, and every fiqh citation says so openly. Scripture is presented as scripture — the Qurʾān and the Ṣaḥīḥayn belong to everyone. A household setting records your school of thought, and Shāfiʿī, Mālikī and Ḥanbalī texts are on the roadmap.
Not yet, and Duha won’t pretend otherwise. Today the corpus is machine-verified: a validator proves every Arabic character is a byte-verbatim slice of its canonical source, on every build. A scholar review is underway, and until a named scholar signs off, every answer carries its true verification status.
It stays on Duha’s own self-hosted server: never sold, never shared, never used to train a model. Inference runs on open-weight models inside confidential-computing enclaves, so even the compute provider can’t read your questions. No ads, no analytics scripts — not even on this page.
Yes — child-safety is enforced on the server, not hidden in a settings menu. Agent tools are locked out for child accounts, companion-style intimacy is refused warmly with a redirect to homework and stories, and crisis routing preempts everything else, always.
Sign-ups open in waves while the scholar review and the corpus work continue. Leave your email below and we’ll write when the window opens, إن شاء الله.