Privacy policy

Your data, on your device and beyond.

Updated2026-10-15

Local data

This policy covers the Anan iOS app and anan.sh. Conversations, memories, character instructions, skills, submitted attachments, task records, workspace files, usage records and settings are saved on your device to provide the features you use. Data comes from your input/imports, service responses and tools you use. Credentials are stored in the system Keychain; app files may enter device backups according to system settings.

Anan does not require an Anan account or use an Anan-operated server to relay model requests. However, model requests leave your device. Local-first and using your own API key do not mean there is no third-party sharing.

What is sent to AI

To answer you or complete a task, model requests may include your messages, relevant history or summaries, character/loaded skill instructions, memory included in context, submitted images/text files, and tool calls/results. This also applies to summaries, context rebuilding and scheduled tasks. The receiving service processes authentication and request metadata such as IP address, model and usage; account-based requests may include account/session identifiers.

Anan does not automatically upload all memories or all device data. A profile may be included when context starts; other memory bodies are included when read or selected for context. Existing context can be sent again, including to a new recipient if you switch services.

Screenshots and sensitive data

When you use supported tools and grant the relevant system permissions, requested contact, calendar, reminder, read-only Health or location results can enter the conversation and later model requests. Screen observation requires full-screen sharing on a supported iOS version; the captured image is saved with the conversation and can be sent to the model. It may show personal information from other apps. This is not an upload of every record on your device.

Submitted attachments may be sent to the model. Image preparation removes metadata but not visible personal details. Voice recognition currently runs on-device without saving or uploading recordings; submitted recognized text is handled like other messages.

Who receives it

Requests go to the actual configured address. Presets include OpenAI, Anthropic, Google Gemini, DeepSeek, Moonshot/Kimi, Zhipu, Z.AI, Alibaba Cloud/Qwen, Xiaomi MiMo, MiniMax and OpenRouter. Regional services may have different operators and terms. Preset hosts and official references are listed below.

A connection name is only a label. An OpenAI-compatible endpoint is not necessarily operated by OpenAI. Check the actual address and operator. Proxies and OpenRouter may forward content to downstream providers, whose terms also apply; the app cannot identify every downstream operator from a custom URL alone.

Official policies and references

Other external services

Enabled MCP services receive authentication/connection requests and invoked tool arguments to perform the requested operation. Web search/page reading sends queries or target URLs to your selected Tavily, Firecrawl, Exa or Parallel service. Remote Sandbox sends relevant commands, code and files to Daytona or E2B. These services also receive credentials and request metadata; results may later be included in model context. Sandbox code may contact additional destinations.

Optional OpenAI/GitHub sign-in contacts those operators for authentication and permitted access. Apple Maps/geocoding/weather may receive tool queries or coordinates. Links, Shortcuts and remote images can contact their respective services. These are separate recipients and purposes, not all AI services.

Official policies and references

Consent and withdrawal

Before starting model inference without a saved consent, Anan shows an AI data-sharing notice with the data categories, purpose, supported services, current connection and its address. Choose Confirm to allow sharing or Cancel to stop that request. This is a single consent covering the AI services and gateways you configure, including services added later; changing a connection does not trigger a new recipient-specific confirmation. A new disclosure version requires confirmation again.

In Me → Withdraw AI sharing consent, you can clear this consent for all model services. Later model requests require confirmation again; background requests without consent are blocked until you return to the app to confirm. Withdrawal cannot recall requests already sent or delete data held by a recipient. System permissions, API keys/sign-in and tool approvals are separate from AI sharing consent. Reading this policy does not authorize sharing.

Third-party protection

Retention, training use, processing location and security depend on the actual service, account, region and contract. OpenAI/Anthropic API terms have retention exceptions; Gemini distinguishes paid and unpaid services, with restrictions on personal/sensitive content under some unpaid terms. A proxy or consumer/account plan must not be assumed to have the same rules as an official API. See the official references.

Applicable data-processing agreements can provide confidentiality, security and deletion safeguards. Equal protection has not been confirmed for every preset, plan, downstream service or custom endpoint. Consent does not itself verify a service’s safeguards. In particular, confirm the operator and applicable protections of a custom endpoint or gateway before sharing personal content. We do not promise that every provider uses no data for training or retains nothing.

Official policies and references

Retention, deletion and your choices

Local conversations and memories remain until you delete or reset them. Deleting a conversation does not clear all memories, skills, tasks, settings or usage records. Check that deletion/reset completed. Keychain credentials may survive uninstalling; remove them or revoke them with the provider when needed. Manage device backups separately.

Third-party copies follow each recipient’s retention and legal rules. Use its console, privacy contact or remote-workspace deletion controls to request removal. Disabling a service, stopping compute, clearing local history or withdrawing consent does not delete previously sent prompts, remote files, snapshots, logs or backups. Stop tasks/sharing and disable connections or system permissions to prevent further use; saved results can remain in existing context.

Security and support

Anan uses iOS storage protections and Keychain; model credentials do not sync through Keychain. Model requests use HTTPS and reject redirects. MCP also supports plain HTTP, which lacks TLS confidentiality. No method guarantees absolute security. The app has no advertising, cross-app tracking or third-party analytics SDK. Local diagnostics, including explicitly enabled debug request capture, may contain data.

If you email support, we receive the contact details and material you provide to investigate and reply. Remove secrets first. For privacy questions or access, correction and deletion requests, contact [email protected]. Contact us about retention or deletion of support material. We cannot delete provider-held copies on your behalf.

The website and updates

anan.sh uses anan-language to remember your language for up to one year; you can delete it in your browser. Existing self-hosted Plausible at plausible.haoran.cv receives page/referrer/outbound URLs for visit statistics, without analytics cookies or access to app data. Hosting/analytics infrastructure also processes IP addresses, user agent and technical request logs; retention depends on the infrastructure and log settings. Contact us with retention or deletion questions.

We update the policy and date when practices change. App Store labels must reflect Apple’s definition of collection and applicable third-party retention; no Anan backend does not imply Data Not Collected.

Official policies and references