Avyra

Terms of Service

The terms that govern your use of Avyra.

Terms & Conditions

Last updated: 21 May 2026

Please read these Terms & Conditions ("Terms") carefully before using Avyra ("the App"). By installing or using the App you agree to be bound by these Terms.

1. About the App

Avyra is an on-device AI assistant. It runs language models, speech recognition, and text-to-speech locally on your iPhone or iPad. The App may optionally make outbound web requests when you enable tools that perform them (the HTTP tool, a user-installed plugin with the http capability, a workflow step pointed at a server LLM provider, or a workflow step pointed at an MCP server). Each of those network destinations is reached only because you configured it.

2. Acceptable Use

You agree to use Avyra only for lawful purposes and in a manner consistent with the acceptable-use policies of any underlying AI models you select (see Section 7) and any third-party services you connect (see Section 6). You will not:

  • Use the App to generate, store, or distribute content that is illegal, harmful, abusive, defamatory, or that infringes the rights of others.
  • Attempt to reverse-engineer, decompile, or extract the source code of the App beyond what is permitted by the open-source components disclosed in Licenses.
  • Bypass safety filters or model usage policies declared by the model providers or third-party services referenced below.

3. User-Installed Plugins

Avyra supports user-installed JavaScript tool plugins ("Plugins"). Plugins run in a sandboxed JavaScriptCore context with a limited capability surface declared in each Plugin's manifest. You are responsible for any Plugins you choose to install. Avyra does not vet third-party Plugins and provides no warranty as to their behavior. Capabilities granted to a Plugin (e.g. http, clipboard, share, secrets) take effect immediately on install and persist until you uninstall the Plugin.

4. Workflows

Avyra includes a workflow builder that lets you assemble multi-step automations from LLM steps, native capability calls (Calendar, Reminders, HealthKit, Location, Files, Secrets), JavaScript transforms, plugin tool invocations, branches, loops, and outputs. Workflows can be triggered manually inside the App, via the system Shortcuts app (as AppIntents), via Siri, or via the avyra://run?workflow=<id> URL scheme.

You are responsible for the workflows you author or import. Workflows execute the steps you wired into them; the App makes no decisions about which capabilities a step calls or which data it sends where. Workflows triggered in background contexts (Shortcuts, Siri, AppIntents) do not have access to server LLM providers or any other network destination — those steps will fail closed.

5. Credentials & the Credential Vault

The credential vault stores secrets used by workflow steps and plugins to talk to outside services (server LLM tokens, MCP server credentials, plugin-declared secrets entries). Metadata is stored in UserDefaults; the secret payload is stored in the iOS Keychain, keyed by credential id. You can mark individual credentials as biometric-gated (Face ID / Touch ID required at read time).

You are responsible for the credentials you save. Avyra does not validate them against the remote service before saving. Rotating, revoking, or deleting credentials is your responsibility; Avyra provides a UI to do so but does not enforce a rotation schedule.

6. Server LLM Providers & MCP Servers

Workflows may be configured to route individual LLM steps through a server-side LLM provider — OpenAI, Anthropic, or Google Gemini — or to call an MCP (Model Context Protocol) server. When you opt into these:

  • Your prompts leave your device. The vendor or MCP operator receives the prompt text, your selected model, your sampling parameters, and any other fields the request shape requires.
  • The vendor's or operator's terms govern that traffic. You must comply with each service's own terms of service, usage policy, and data-handling policy. Notable touchpoints:
  • Avyra is not a reseller or intermediary. Your billing relationship is directly with the vendor (you supply your own API key). We do not see your token, your usage, or your invoices.
  • Background contexts cannot use these. Workflows invoked via Shortcuts, Siri, or background AppIntents do not have network access for server LLM steps; assign on-device Apple Intelligence to any step you want to invoke from those contexts.

7. Third-Party On-Device Model Licenses & Acceptable Use

Avyra lets you download and run open-weight language models published by third-party organisations. Each model is licensed and governed by its own terms. You are responsible for complying with the terms of any model you choose to use. The following models may be available via the in-app model picker:

Google Gemma (Gemma 2, Gemma 4)

Gemma models are licensed by Google under the Gemma Terms of Use and subject to the Gemma Prohibited Use Policy. Notable prohibitions: you may not use Gemma to perform, facilitate, or generate output for activities that violate applicable laws or harm individuals — including the generation of child sexual abuse material, content promoting violence, terrorism, illicit drug use, or sexually-explicit deepfakes of real people; the development of malicious software or weapons; or activities that intentionally deceive (e.g. fraud, impersonation). Refer to Google's policy for the full list. You may use, modify, distribute, and fine-tune Gemma subject to compliance with these terms.

Meta Llama (Llama 3.x)

Llama models are licensed under the relevant Llama Community License (e.g. Llama 3 Community License) and subject to Meta's Acceptable Use Policy. Notable prohibitions: you may not use Llama to violate any applicable law or rights of others; to engage in or promote violence, harassment, or illegal activity; to generate CSAM; to impersonate others in a deceptive manner; to engage in unauthorized practice of medicine, law, or other regulated professions; or to develop weapons of mass destruction or critical-infrastructure attacks. Commercial use is permitted subject to a monthly active user threshold defined in the license; consult the full license for details.

Alibaba Qwen (Qwen 2.5, Qwen 3.5)

Qwen models are typically licensed under either the Tongyi Qianwen License Agreement or the Apache 2.0 license, depending on the specific model variant. Smaller Qwen variants (including those bundled in the in-app catalog) are generally Apache 2.0; larger variants and Qwen-Max are under the Tongyi Qianwen license, which carries usage restrictions for users with more than 100 million monthly active users. Notable prohibitions across both licenses: you may not use Qwen for activities prohibited by applicable law; to harm individuals or groups; to generate CSAM or illegal content; or to mislead users about the source of generated content. Consult the license text for the specific Qwen variant you download for the authoritative terms.

Plain-English summary Don't do anything illegal, harmful, or deceptive with any of these models — the same things that would get you in trouble using ChatGPT or Claude get you in trouble here. When you point an LLM step at a server provider, that vendor's rules apply on top of these Terms. When you bring your own MCP server, that operator's rules apply too. You agreed to each of those policies when you supplied your key or URL.

8. No Warranty

The App is provided "as is" and "as available," without warranty of any kind, express or implied. Model outputs may be inaccurate, incomplete, or offensive. You should independently verify any information generated by the App before relying on it, particularly for legal, medical, financial, or safety-critical decisions.

9. Limitation of Liability

To the maximum extent permitted by law, the operators of the App will not be liable for any indirect, incidental, special, consequential, or punitive damages arising from your use of the App, any Plugin you install, any workflow you author, any server LLM provider you configure, or any MCP server you connect.

10. Open-Source Components

Avyra includes open-source software listed in the Licenses screen. The corresponding license text is reproduced there. Your use of those components is governed by their respective licenses.

11. Changes to These Terms

These Terms may be updated when new features ship or third-party policies change. The "Last updated" date at the top of this document reflects the current version. Material changes will be highlighted in the FAQ or release notes.