Premium recipes you can trust — every amount measured, every step cued, then rewritten for your pantry, your allergies, and the time you've got. Plus voice-first cooking when your hands are full of dough.
Pantry-aware ways to begin
Most recipes fail you in the same place: they assume. They say “season to taste” and leave you guessing, “cook until done” and never say what done looks like, or call for a pan you don’t own. Saffra’s catalog is written the other way round — every ingredient with an amount and how to prep it, the kit listed before you start, and a step that tells you what to look for. Then she rewrites it for you.
She checks what you have before suggesting anything. If the cilantro's going yellow, that's tonight's dinner.
Messy hands? Say "hey saffra, next step" and she narrates. Set timers by voice — they live in the Dynamic Island.
Every recipe is rewritten for your pantry, equipment, allergies, and skill level. Beginners get sensory cues. Pros get the chef move.
Declared allergies are treated as non-negotiable. Saffra names the substitution or passes on the recipe rather than quietly proceeding.
Amounts, prep notes, the equipment you need and a cue for what done looks like — with allergen flags on every recipe.
Substitutions named out loud — sensory cues for beginners, the chef move for pros.
Open Saffra at 7 AM and the Recipes tab is a breakfast app. Open it at midnight and it's a late-night one. The quick rail, the order of every section, and Saffra's headline pick all re-rank around the meal you're actually about to cook — no tapping, no filters, no settings.
"Quick breakfast" leads. Pancakes and shakshuka up front — heavy dinner mains stay out of the morning rail.
"Quick lunch" takes over — bowls, salads, and fast mains, led by Saffra's lunch pick.
"Saffra picks for tonight." The dinner canon — pasta, chicken, soups — ranks to the top for the main event.
"Late-night quick" — nothing over 25 minutes, nothing that needs a supermarket run. Fried rice territory.
Generic "AI cooking" tools paste a recipe into a chatbot and call it a day. Every layer of Saffra is built for cooking specifically — with knowledge, memory, and rules a general-purpose AI doesn't have.
When you ask "what can I make tonight?", Saffra isn't answering cold. Before she says a word, she can check your pantry, your cook history, the clock, and your profile — the same way a chef glances at the walk-in before calling the special. Watch the tool chips in a real chat: that's her actually looking, not pretending to.
Powered by advanced AI — shaped into a real chef. Saffra has a name, a voice, signature moves, and hard safety rules. Her tone calibrates to your skill: beginners get sensory cues ("when oil shimmers, when it smells nutty"), pros get terse technique-first ("hit it with a nappé"). One model, infinite cooks.
No Saffra recipe is improvised the moment you ask. The catalog is written in advance and reviewed in-house, with allergen flags on every recipe — and it is written to be cooked from: 99.8% of ingredients carry a measured amount, 97.2% say how to prep it, every recipe lists the equipment it needs, and 99.3% name a cue for what done looks like. Saffra's catalog: 2,300+ recipes across 25+ cuisines, and every one of them is semantically searchable (vector embeddings). Then she adapts it to your pantry, your skill, your equipment.
Rolling conversation summaries. Cook history with star ratings. Live pantry awareness. Recent-suggestion tracking so she doesn't pitch the same dish twice. The longer you cook with her, the sharper she gets — leaning into what you loved, steering around what flopped, varying within genres you've succeeded at, never repeating last Tuesday's dinner unless you ask.
Saffra doesn't just talk — she works. She has nine purpose-built tools: pantry search, recipe lookup, recipe adaptation, ingredient substitution, technique explainer, voice timers, unit conversion, recipe saving, profile updates. She picks the right tools for each turn and chains them — "what's for dinner?" might fire pantry-search → recipe-lookup → adapt-for-user, all silently before answering.
Declared allergies are non-negotiable — hard refusal rules, not soft suggestions. She'll either swap loud and clear or pass on the recipe entirely. Cross-contamination flags for serious allergies (peanut, tree nut, shellfish, sesame, gluten-for-celiac). USDA-safe internal temperatures cited in every recipe that needs one.
Streaming neural TTS — Saffra's signature voice — served from a content-addressed cache so the same line never re-renders twice. Speech recognition uses Apple's Speech framework, which is server-backed — audio goes to Apple, not to us. Every question you speak in Cooking Mode is also re-transcribed by OpenAI's Whisper for accuracy, then discarded. The wake phrase "hey saffra" is only armed inside Cooking Mode, never app-wide. Voice timers land in Live Activities, including the Dynamic Island.
An amount, a unit, and how to prep it
The kit you need, before you start
What done looks like, not just how long
Allergen flags on every recipe — then she rewrites it for your pantry, your skill, your equipment.
Tap "Let's cook" on any recipe and the whole phone becomes a chef at your shoulder. One step per screen, big enough to read from across the counter, with a photo of the dish, the exact ingredients this step needs, and a timer she's already sized for you.
Saffra narrates in her signature voice, then listens. "Hey saffra, next step" — "replay that" — "how long on the timer?" Your hands stay in the food.
Each step card pulls out exactly what you're reaching for — "spaghetti, salt" — so you're never re-scanning the full list mid-sizzle.
"15 min — saffra suggests this timer." One tap (or one sentence) and it's running in the Dynamic Island, even if you leave the app.
"Why reserve the pasta water?" The ask-saffra bar answers inline without leaving the step — and a "coming up" preview keeps you one move ahead.
Any recipe, one tap. The screen clears to a single step with a progress bar — no scrolling a wall of instructions with wet hands.
Saffra narrates the step and waits. The wake phrase "hey saffra" is armed only here — never anywhere else in the app.
Say "set the timer" and it lands in Live Activities — Lock Screen and Dynamic Island. Multiple timers, all labeled.
The last step ends with "I'm done — tap to snap a photo." Your cook lands in your history, and she remembers how it went.
The Saffra tab opens with a time-of-day greeting, your pantry at a glance, and quick-starts built from what's actually in your kitchen. Ask anything — she answers like a chef, and when a recipe is the answer, she hands you a real card from the recipe catalog: one tap to open it, one more to cook it.
"What can I make tonight?" gets a recipe card pulled from the prepared catalog — never a list of steps invented on the spot.
Tool chips show each check as it happens — including pantry scan and catalog search — so you can see what informed the answer.
Morning greeting, evening greeting, late-night greeting — with one-tap prompts shaped by the hour and your on-hand ingredients.
Saffra can only cook from your kitchen if she knows what's in it — so getting food into the app takes seconds, not typing. And once it's in, freshness is tracked: things about to turn get flagged and pitched for tonight's dinner instead of the trash.
Photograph a shelf and she identifies what's on it. Items land categorized, no typing. This one photo is sent to our server to be read; receipts and cookbook pages are read on your phone.
On-device OCR reads the lines, AI parses them into clean items — your photo never leaves the phone.
Lookup runs against Open Food Facts — items arrive pre-tagged with category, expiration, and allergen flags.
Items going stale get flagged in the pantry and surfaced in chat — "the cilantro's going yellow, let's use it tonight."
Search runs on vector embeddings across every recipe in the catalog — so "cozy noodle soup" and "something bright with lemon" work as well as exact titles. Every result comes back as a photo card with time and cuisine up front, and refinements for time, diet, and cuisine are one tap away.
Semantic search understands the dish you're describing — no more guessing the magic word a recipe site indexed.
Under 30 minutes. Vegetarian. Thai. Stack refinements and the results re-rank instantly.
A recipe page should answer one question fast: can I cook this, tonight, with what I have? Timing, difficulty, and servings sit at the top; ingredients are checked against your pantry; and the whole thing ends in one button — "Let's cook."
Prep, cook, and total time, difficulty, and servings you can scale — the recipe recalculates quantities for you.
What you have is marked; what you're missing can go straight to a shopping list. No mid-recipe surprises.
Calories and macros per serving, plus allergen callouts that persist through every adaptation Saffra makes.
One tap rewrites the recipe for your allergies, equipment, and skill — swaps named loud and clear, never silent.
Each feature exists for one reason: get you from "what should I cook" to dinner on the plate.
She suggests recipes that use what you already have. Items going off get used first. Less waste, fewer trips.
"Hey saffra, next step." She listens the whole time you're cooking. Voice timers land in the Dynamic Island.
Every recipe rewritten for your pantry, equipment, allergies, and skill — beginners get sensory cues, pros get the chef move.
Paste a URL, snap a cookbook page — Saffra imports it and adapts it for you. Build a personal library that learns your taste.
Scan groceries straight into your pantry — items land with categories, expiration dates, and allergen flags pre-filled.
Declared allergies are treated as hard rules. Saffra names substitutions clearly or passes on the recipe rather than quietly proceeding.
Tell her your level once. Beginners get every technique explained with sensory cues. Pros get terse, technique-first direction.
"Set rice for eighteen minutes." It's running. Multiple timers track on Lock Screen and Dynamic Island so you don't miss a beat.
One-minute conversational onboarding. Diet, allergies, equipment, skill, the cuisines you love.
She pulls from the prepared catalog, checks your pantry and your time, and names the dish — confidently, adapted to you.
Voice-first Cooking Mode reads you each step. Set timers by voice. Ask "why does this matter?" any time.
Rate the result. She remembers. Future suggestions lean into what worked, and quietly skip what didn't.
Start free. Upgrade when you're ready to cook with her every day.
Try the chef
For everyday cooks
Three differences matter. First: a general model drafts a recipe the moment you ask, from model knowledge alone. Saffra's flagship recipes are written in advance and complete enough to cook from — 99.8% of ingredients carry a measured amount, 97.2% say how to prep it, every recipe lists the equipment it needs, and 99.3% name a cue for what done looks like. Second: the assistant works with your pantry, equipment, allergies, skill level, and current cooking step rather than treating each question in isolation. Third: product-specific tools handle catalog search, substitutions, techniques, and hands-free cooking, so the model is one part of a larger cooking system.
The core assistant uses a multimodal AI model, with a faster utility model for tasks such as parsing receipt text and folding older conversations into summaries. Recipe search uses vector embeddings over the curated library. Voice synthesis uses a neural TTS voice with content-addressed caching, while speech recognition uses Apple's Speech framework, which is server-backed — the audio reaches Apple, not us. In Cooking Mode the clip is also re-transcribed by OpenAI's Whisper, for every spoken question rather than only noisy ones, then discarded. Receipt and cookbook OCR runs on-device through Apple's Vision framework; shelf photos are sent to the server for vision analysis and are not stored. The recipe catalog is written ahead of time rather than improvised per request, and every recipe carries allergen flags.
Two things keep her grounded. One: when you ask for a specific dish or "what should I cook?", she calls a tool that searches Saffra's own recipe catalog — every recipe in it written in advance — instead of improvising one from memory. So you get a real recipe, adapted for you, not a hallucination. Two: for techniques and substitutions she calls dedicated tools (technique explainer, substitution lookup) rather than guessing. She also won't claim to have personally tasted something — she'll say "by the numbers, this should read bright and savory" instead of "I love this." If she's not sure, she says so.
Saffra works across cuisines including Thai, Mexican regional, Korean, Vietnamese, French, Italian regional, Chinese regional, West African, Levantine, and Peruvian; dietary needs including vegan, keto, halal, kosher, FODMAP, Whole30, and celiac; skill levels from first-time cooks to technique-focused home cooks; and kitchens ranging from induction-only or no-oven setups to fully equipped spaces. For dishes outside the flagship catalog, it applies regional and technique-specific guidance and states uncertainty when appropriate.
Yes. Conversations are persisted, your pantry is tracked, and recipes you rate 4-5 stars influence future suggestions. After about 24 turns the older messages fold into a compact summary so context stays sharp without bloating. She knows what you've cooked recently (and won't repeat last Tuesday's dinner unless you ask), what you've loved, and what fell flat. She'll weight suggestions toward your wins, vary within genres you've succeeded at, and treat anything she's already pitched in this thread as off the table.
You tell her — but quickly. Snap a grocery receipt and on-device OCR extracts the lines, then a lightweight AI model parses them into clean line items (your photo never leaves the phone). Or scan a barcode — the lookup runs against Open Food Facts and items land in your pantry pre-tagged with category, expiration, and allergen flags. Or photograph a shelf and let her name what she sees — that photo is the one that does leave, going to our server to be read by a vision model before you confirm the list (it isn't stored). Or just type things in. As you cook, she tracks what's used and what's running low. Items going off ("the cilantro's gone yellow") get surfaced for tonight's dinner instead of going in the trash.
Never. Text-mode chat is the default and works exactly the same. Voice mode is for when your hands are deep in dough or holding a knife. Toggle it per-session in Settings, or open Cooking Mode for full hands-free narration with the wake phrase "hey saffra" — note that the wake phrase is only armed inside Cooking Mode, never app-wide, so the mic isn't listening when you're not cooking. Voice timers fire silently and land in Live Activities (Lock Screen + Dynamic Island). Free tier is text-only; Pro unlocks voice cooking.
Declared allergens are treated as hard product rules. Saffra either names a substitution or declines the recipe and offers an alternative rather than silently proceeding. For serious allergies (peanut, tree nut, shellfish, sesame, and gluten for celiac users), it also flags cross-contamination risks in shared oils, fryers, and cutting boards. Adapted recipes carry a persistent allergen notice, and recipes include target internal temperatures where appropriate. Saffra is still an AI system: verify allergens and cooking temperatures yourself, and do not treat it as medical or dietary advice.
Free gives you 10 chef messages in total — a lifetime allowance, not a daily one — plus full recipe-library browsing, one recipe adaptation, pantry tracking, and the same safety rules. Pro lifts the free tier's 10-message chef-chat cap (fair-use rate limits still apply), and adds voice cooking, ongoing recipe adaptation, receipt and barcode scanning, recipe import from URLs and cookbook photos, and a personal library that learns from your cooking history.
Saffra separates recipe writing from real-time adaptation. The flagship catalog is drafted in advance and reviewed in-house, with allergen flags on every recipe — it is not improvised the moment you ask. What that buys you is completeness: a measured amount on 99.8% of ingredients, a prep note on 97.2%, an equipment list on every recipe, and a sensory cue for the finish on 99.3% (measured 2026-08-19 across the shipped catalog). When you ask for the dish, Saffra starts from that prepared catalog entry and adapts it to your pantry, equipment, allergies, and skill level. The adaptation is generated by an AI model; the catalog entry is not.
Saffra is out now on the App Store for iPhone. The first release is focused on iPhone; no timing has been announced for other platforms.
Sign in with Apple or Google. We never sell data. We never use your conversations to train models. We never track you across apps or websites. When you use voice mode, audio is transcribed by Apple's Speech framework, which is server-backed — the audio reaches Apple, not us. In Cooking Mode the clip is also sent to OpenAI's Whisper for a second, more accurate transcription — every spoken question, not only the noisy ones — and discarded afterwards. Receipt and cookbook photos run through Apple's Vision framework on your phone — only the extracted text leaves the device. The one image that does leave is the shelf photo you take to stock your pantry: that is sent to our server and read by OpenAI's vision model, and not stored. You can delete your account and every byte of your data from Profile / Me at any time. Full policy at /privacy.
Premium recipes, complete enough to cook from. Adapted to your kitchen. Cooked with you, hands-free. Out now on the App Store. iOS 17 or later.
iOS 17 or later. Free to download.