Privacy at Relay

What product usage events Relay collects

After you allow analytics, Relay records defined events such as a product page view, login attempt or completion, question submission, first answer completion, agent checks, synthesis completion, feedback, saving, sharing, and a return session. Events include random visitor and session identifiers, Relay’s internal user ID when signed in, the page path, sanitized campaign fields, referrer domain, related thread or request IDs when available, and tightly limited event properties.

Relay stores the first-party event log in its own application database. It is the source of truth for product metrics. Duplicate event IDs are ignored. Each event also carries an event timestamp and schema version so retries and future migrations can be handled safely. Relay does not use fingerprinting or collect device fingerprint attributes.

Authentication and account data

Relay uses Supabase Auth for email/password, Apple, Google, provider-gated Microsoft sign-in, and provider-gated WeChat sign-in. Relay uses your internal account ID to associate your library and public contributions. Your email, when present, is used for account access and displayed in your account screen; it is not sent to product analytics. Email/password requests use a same-origin Relay transport endpoint and are forwarded only to Supabase Auth. Relay does not log or retain those request bodies, and passwords are never written to Relay’s product-event, model-call, or application databases. Relay never receives your Apple, Google, Microsoft, or WeChat password.

When WeChat sign-in is enabled, Relay’s server exchanges the one-time WeChat authorization code and briefly holds the resulting WeChat account identifier, nickname, and avatar URL only to complete the Supabase OAuth handoff. WeChat access and refresh tokens are not stored in Relay’s application database. The short-lived handoff records expire within ten minutes; Supabase then keeps the linked provider identity needed for future sign-in.

Questions, answers, and AI providers

In a Relay website Battle, your question and a bounded amount of recent conversation context are sent through Relay’s server to the two models you select from ChatGPT (OpenAI), Gemini (Google), DeepSeek, Grok (xAI), and Kimi (Moonshot AI), so each can produce an independent answer. ChatGPT and Kimi remain the default pair. The question, both model identities, and both answers are then sent to the Judge model you select from the same available provider list; Gemini remains the default Judge. The Judge may be one of the two contestants if you explicitly choose it. An imported single-answer check uses its selected review provider and is not automatically sent to every Battle provider.

Relay does not put prompt or response text in analytics or model operational logs. When you are signed in, each completed question, answer, and Relay evaluation is saved to your private My Questions history. Anonymous and incomplete requests are not intentionally persisted by Relay as account content. A private result becomes a public Community post only when you explicitly choose to post it.

When the optional Council Review experiment is available and you choose “Challenge both answers,” each selected answer model receives the question and the other model’s anonymous answer for review. The selected Judge then receives the question, both anonymous answers, and both critiques to produce a combined answer. Relay stores the original answers, critiques, combined answer, status, failures, usage, cost, and optional Yes/No feedback as private, owner-scoped Council state. It is not substituted for the original Battle. Saving keeps it private; it becomes public only if you explicitly share a snapshot that includes it.

AI providers process submitted content under their own terms. Relay requests non-persistent processing where the provider interface supports it, but does not make broader promises about a provider’s systems. Kimi’s international OpenPlatform terms identify Moonshot AI PTE. LTD. as the API provider. Review Kimi’s privacy policy and platform terms. DeepSeek’s Open Platform is operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd.; review its privacy policy and Open Platform terms before submitting content you consider sensitive or confidential. Grok’s API is operated by X.AI LLC. xAI says API inputs and outputs are not used to train its models without explicit permission and are retained for up to 30 days for abuse and misuse auditing before deletion; review xAI’s API security FAQ and enterprise terms.

If you connect an AI client to Relay, the client can submit a question only after you approve the connection. A plain Relay question is sent to Kimi K3 for an independent answer, and Kimi does not receive the existing ChatGPT answer. An explicit Battle can use an existing ChatGPT answer or have Relay request one from OpenAI, while Kimi answers the question independently. Only an explicit compare, judge, or “who is right?” request sends the question and both Battle answers to Google’s Gemini model. Each Kimi answer or Battle is saved as a private thread in your Relay library. You can deny the connection without changing your account.

Model-call operational metadata

For reliability and cost control, Relay records request and thread IDs, internal user ID when signed in, provider, model, task type, routing, token counts returned by the provider, latency, estimated cost when a verified price is configured, success or failure, normalized error code, retry count, and timestamp. These records do not contain prompts or generated answers. Estimated costs are estimates, not invoices.

Free usage and complimentary Pro

Relay currently does not collect payment information. For signed-in accounts, Relay totals the estimated model-call cost described above within each UTC calendar month to operate the free usage allowance. If you apply by email for complimentary Pro, your email is handled by the email provider you use and Relay stores only your internal account ID and complimentary access status needed to grant or revoke access.

Feedback

The two structured feedback choices and any optional text you submit are stored by Relay for product improvement. Optional feedback text is limited to 1,000 characters, rate-limited, and never forwarded to a third-party analytics provider. Feedback is optional and does not block continued use.

Third-party analytics

Relay supports an optional server-side PostHog connection. It is used only when a PostHog project is configured and you have allowed analytics. Only the same sanitized event taxonomy is forwarded. Relay does not use PostHog autocapture or session replay, and never forwards form contents, page text, emails, names, questions, answers, or authentication data.

Cookies and browser storage

Relay uses browser storage for your language, analytics preference, random visitor and session identifiers, temporary in-progress drafts, and Supabase authentication session. The browser session identifier is also used for anonymous usage limits and abuse prevention, even when optional analytics is declined. A random visitor identifier and your consent choice and timestamp remain on this device so Relay can honor that choice; declining clears campaign attribution and product-funnel progress while keeping authentication and security controls working.

Campaign and referral data

Relay accepts only utm_source, utm_medium, utm_campaign, and utm_content. Values are length-limited and character-validated. Relay stores only a referrer domain, never a full referrer URL that could contain sensitive query parameters.

Retention

Default retention is 30 days for anonymous product events, 180 days for authenticated product events, 90 days for model operational logs, and 180 days for optional feedback. Anonymous Council Reviews are retained for 30 days and authenticated Council Reviews for 180 days by default. These periods are configurable. Relay runs a daily cleanup procedure and provides an owner-only manual cleanup action. User-created threads are not removed by analytics cleanup. Complimentary-access records are kept while needed to operate the account.

Deletion requests

Automated account deletion is not yet available in the web product. To request deletion of your Relay account, stored threads, contributions, complimentary-access record, Council Reviews, feedback, or associated analytics records, email relay.ai.info2026@gmail.com. Relay will confirm the account and scope before taking action.