Updated October 19, 2026
AI Processors
nagu uses AI processors only to provide or improve user-facing inbox features for the connected account. This page is generated from the same runtime manifest that routes live AI traffic, so the list below matches the processors actually in use.
Retention for AI requests
Language-model requests go to OpenAI's API directly. OpenAI does not train on API data. It keeps prompts, message excerpts, and outputs for up to 30 days to monitor for abuse, then deletes them.
If the direct route is not configured, requests go through the Vercel AI Gateway instead, with the same no-training promise.
We also run shadow checks. nagu sends an incoming message's excerpt to TypeSafe, a second decision model, to compare its triage answers with ours. Shadow checks run on incoming mail only. We record the results and never act on them. TypeSafe keeps requests for as long as it needs to run its service and publishes no fixed period.
Search by meaning runs on Voyage AI Innovations, Inc. For every message you send or receive, nagu sends Voyage the subject and the first 8,000 characters of the body. It also sends thread digests, contacts, agent summaries and your search queries. Voyage may use the mail it receives to train and improve its models. It keeps requests for as long as it needs to run its service and publishes no fixed period. Its terms are at voyageai.com/privacy.
Three separate statements about training
- No general model training by us. nagu does not train its own general-purpose AI models on your mail.
- Per-user personalization only. nagu learns your voice and habits to power your own features — a per-user, account-scoped model that is yours alone, not a shared one, and you can clear it at any time.
- What providers do. OpenAI does not train on data sent through its API, every fallback provider is contractually blocked from training on your content, and TypeSafe states in its privacy policy that it does not train on the inputs it receives. Voyage, which builds the search index, may use the mail it receives to train and improve its models. OpenAI keeps API data for up to 30 days for abuse monitoring.
These are three different statements. The first is about our own models; the second is about a feature that belongs to you; the third is what each provider below does with what it receives. None of them implies the others.
Current AI processors
This table is generated from nagu's runtime compliance manifest — the same source that decides where AI requests are routed — so it stays in sync with the providers actually in use. OpenAI is called directly; the gateway and Anthropic are fallbacks; TypeSafe runs shadow checks only; Voyage builds the search index. Every route except Voyage carries a no-training posture.
| Processor | Role | Region | Retention and training |
|---|---|---|---|
| OpenAI | Language-model processing for triage, briefs, summaries, reply drafts, and classification decisions, called directly. | United States | Kept by OpenAI for up to 30 days for abuse monitoring, then deleted. No training on your content. |
| Vercel AI Gateway | Fallback routing only, used when the direct OpenAI route is not configured. Not a model-training destination for your content. | United States | Zero retention: processed for the request, then dropped. No training on your content. |
| Anthropic | Language-model processing only when a request is routed to Anthropic through the gateway fallback. | United States | Zero retention: processed for the request, then dropped. No training on your content. |
| TypeSafe | Shadow checks only: its answers are recorded to compare with our own and never change what happens to your mail. | United States | Kept by TypeSafe for as long as it needs to run the service. It publishes no fixed period, and zero retention is an enterprise arrangement. No training on your content. |
| Voyage AI Innovations, Inc. | Search by meaning. It receives the text of mail we index and your search queries, and may use them to train and improve its models. | United States | Kept by Voyage for as long as it needs to run the service. It publishes no fixed period. It may use this content to train and improve its models. |
Data minimization
nagu sends only the account-scoped excerpts, metadata, instructions, and context needed for the requested AI feature. Two things are broader. Search indexing sends Voyage the text of every message you send or receive, so search can work by meaning. Shadow checks are not a feature you request, and they receive the text of incoming messages so we can compare decision models. We do not send an entire mailbox when a single thread, summary, draft, or brief can be produced from a narrower context.
Changes
Because this page reads from the runtime manifest, the list updates when the providers in use change. Material changes to AI processing are also reflected in the Privacy Policy and the how your mail data moves overview.