Frequently Asked Questions
What is Avyra?
Avyra is an on-device AI assistant for iPhone and iPad. It runs language models, voice recognition, and text-to-speech locally — no servers, no accounts, no analytics. You can chat by text, speak via voice mode, attach images, extend the assistant with tools and user-installable JavaScript plugins, and build multi-step workflows you can trigger from Shortcuts, Siri, or a URL.
Does Avyra send my data anywhere?
By default, no. The only outbound traffic is what you trigger yourself: downloading a model, enabling the HTTP tool, installing a plugin that makes web requests, signing in to iCloud, configuring a server LLM provider (OpenAI / Anthropic / Gemini) and pointing a workflow LLM step at it, or connecting a workflow MCP step to an MCP server you supplied. See the Privacy Policy for full detail.
Which models can I use?
Three kinds:
- Apple FoundationModels — Apple's on-device LLM, available on iOS 26+. Selected by default. The only model that works in background Shortcuts / Siri contexts.
- MLX models — open-weight models you can download from the in-app model picker: Qwen 3.5 (0.8B / 4B / 9B), Qwen 2.5 1.5B, Llama 3.2 3B, Gemma 2 2B, Gemma 4 e2b / e4b. Larger models give better answers but need more RAM and storage.
- Server LLM providers — OpenAI, Anthropic, and Google Gemini. You supply your own API key, save it as a credential, and point individual workflow LLM steps at the provider. Per-token billing is between you and the vendor.
How do I switch models?
For chat: Settings → AI → Manage models. Tap a model to download it (one-time, ranges from ~300 MB to ~5 GB), then select it as the active model. Subsequent chats use the new model.
For a workflow LLM step: open the step in the workflow editor, tap the Model row, and pick Apple Intelligence (the default) or one of your configured server providers.
What are workflows?
Workflows are multi-step automations you can author or import. A workflow chains LLM steps, capability calls (Calendar, Reminders, HealthKit, Location, Files, Secrets), JavaScript transforms, plugin tool invocations, branches, loops, and outputs. Settings → Workflows opens the library. Tap a workflow to run it, edit it, or trigger it from Shortcuts or Siri.
How do I trigger a workflow from Siri / Shortcuts?
Each workflow declares its triggers. If Shortcuts is enabled, the workflow shows up as an AppIntent in the Shortcuts app and Siri ("Run Daily Brief"). If URL scheme is enabled, you can run it via avyra://run?workflow=<id> from another app. Manual triggering inside Avyra is always available.
Will my workflow run in the background?
Yes for most steps — but with one constraint: background contexts (Shortcuts, Siri, AppIntents) cannot reach the network. So:
- Steps using Apple Intelligence run fine in background.
- Steps using a server LLM provider (OpenAI, Anthropic, Gemini) or an MCP server will fail closed in background contexts. Use Apple Intelligence for any LLM step you want to invoke via Siri.
- Steps that need biometric-gated credentials can't prompt for Face ID / Touch ID in background — the credential read returns
nilthere.
What HealthKit data can workflows access?
Read-only, and only what iOS grants you. Today's supported reads include: steps, walking + running distance, active and basal energy burned, exercise + stand minutes, heart rate (latest), resting heart rate, heart rate variability, sleep analysis (asleep / in-bed / stages), mindful sessions, water (dietary water), workouts (type, duration, energy), height, body mass, BMI, body fat %, lean body mass, blood pressure, respiratory rate, oxygen saturation, body temperature, menstrual flow, and walking heart rate average. Each Health category is granted individually via the standard iOS permission sheet on first use. Avyra never writes to HealthKit.
Can I use OpenAI / Claude / Gemini?
Yes, on a per-LLM-step basis. Add a credential (Settings → Credentials → New), configure the provider (Settings → AI Providers → New), then open any workflow's LLM step and pick that provider from the Model picker. The prompt for that step goes to the vendor; per-token billing is between you and them. We never see your token or your invoice.
Where do my API keys go?
Into the iOS Keychain, keyed by credential id. The credential metadata (display name, kind, notes) lives in UserDefaults so the picker UI can render the list without unlocking the Keychain, but the token itself never leaves the secure element except when a workflow step needs to send it as an auth header — and even then it's used in the request and immediately discarded from memory. Optional biometric protection (Face ID / Touch ID) can be applied per credential.
Can I connect to an MCP server?
Yes. Workflows include an MCP tool step that opens an HTTPS / WebSocket connection to a server URL you supply and exchanges JSON-RPC messages. If the server requires authentication, you can pick a credential (API token or username + password) saved in the credential vault. The MCP server operator's privacy policy governs anything sent to them.
What do the tools do?
Tools let the AI model do things it can't do on its own — fetch live data, do exact arithmetic, parse JSON, save a fact about you. Avyra ships with six built-in tools:
- Remember facts — saves durable user facts to the memory store.
- Current time — returns the local date and time.
- HTTP request — makes a web request for live data.
- Calculator — evaluates an arithmetic expression exactly.
- JSON path — extracts a field from a JSON document.
- Regex — matches, extracts, or replaces text with a regex.
Each can be toggled in Settings → AI → Tools.
How do plugins work?
Plugins are user-authored tools written in JavaScript. Each plugin is a .aria-tool file — a small JSON document with the plugin's manifest (name, description, input schema, declared capabilities) and the JavaScript source. Install a plugin by tapping Settings → AI → Tools → Install a tool… and picking the file.
Each plugin runs in its own sandboxed JavaScriptCore context with access only to the capabilities its manifest declares (e.g. http, clipboard, share, secrets). A plugin cannot access another plugin's data, read your conversation history, or escape its declared capability set.
Is voice mode private?
Voice mode uses Apple's SFSpeechRecognizer. When on-device recognition is supported for your locale, audio never leaves your device. When it isn't, the recognizer falls back to Apple's server-side path. We never receive or store voice audio. TTS (the assistant speaking back to you) is either Apple's built-in voices or the optionally-downloadable Kokoro engine — both run on-device.
Do conversations sync between my iPhone and iPad?
Not in real time. Settings sync across devices via iCloud Key-Value Storage immediately. Conversations, memories, and workflows are included in iCloud Backup, so they restore on a new device set up from your iCloud Backup, but are not actively synced device-to-device. Credentials sync via iCloud Keychain if you have that enabled in iOS Settings — Apple end-to-end encrypts that store.
Can I delete everything?
Yes. Settings → AI → Memories → Clear removes the memory store. Settings → AI → Storage manages downloaded models. Settings → Credentials, AI Providers, and Workflows each support swipe-to-delete. Uninstalling the App removes all on-device data; if iCloud Backup retains a copy, you can delete that via iOS Settings → Apple ID → iCloud → Manage Storage.
What about accuracy?
AI models can be wrong, especially smaller on-device models. Don't rely on Avyra for legal, medical, financial, or safety-critical decisions without independent verification. Tools like Calculator and HTTP help with factual answers, but the model still decides when to use them and what to ask. The same caveat applies whether you're using Apple Intelligence, an MLX model, or a server LLM provider.
Is there a pro / paid tier?
No. There are no in-app purchases, subscriptions, or tiered features. The full Avyra feature set is available to every user. If you use a server LLM provider, that vendor charges you directly per token — Avyra doesn't take a cut.