Question 1

Can it archive, label, or send mail, or is that really blocked?

Accepted Answer

Really blocked, and not by the app. It requests one scope, gmail.readonly, which Google enforces at the OAuth layer, so it cannot mark read, label, archive, report spam, or send. The only Gmail methods it calls are messages.list, messages.get, threads.list, history.list, and getProfile. Trays, corrections, and Done state live in a local SQLite file you can delete.

Question 2

Will a free Gateway key finish a full inbox sort?

Accepted Answer

Not quickly. Jev is free on the Vercel AI Gateway free tier, but that tier allows 5 calls per 5-minute window, and the defaults are built around it: JEV_CONCURRENCY 1, JEV_BURST 5, and a 310,000 ms drain interval. Credits remove the limit; on the paid tier the README suggests concurrency 10 and burst 1000, where 1,000 emails sort in about a minute for roughly 3 cents.

Question 3

How are urgency and 'was this written by a human' decided?

Accepted Answer

Each message becomes a plain-text block of sender, subject, headers, and trimmed body, then one call to typesafe-ai/jev answers three questions in parallel: which tray as a choice, how urgent as a 1 to 5 score, and whether a human wrote it to you as a boolean. Because the options are enumerated, the model cannot invent a tray that is not in the list, and each tray shows the top two probabilities.

Question 4

Do I need a TypeSafe account to run this?

Accepted Answer

No TypeSafe account directly, since Jev is called through the Vercel AI Gateway, but you need a Vercel account and a vck_ key. The heavier setup is Google Cloud: a project, the Gmail API enabled, an external OAuth consent screen with your address as a test user, and a web client whose redirect URI is http://localhost:3000/api/auth/callback/google. Testing-mode refresh tokens expire weekly.

Question 5

How much mail does it pull, and can I change the fetch pace?

Accepted Answer

The first sync takes your SYNC_LIMIT newest inbox messages, 20 by default and 1000 for the full inbox. Sync new fetches arrivals through the Gmail history API, and Fetch more reaches 100 older messages per click. GMAIL_RATE_PER_MIN defaults to 100 and should only rise after you raise the Gmail API quota in Cloud Console; sorting starts once the first 100 have landed.

Question 6

Can I try the interface before wiring up a mailbox?

Accepted Answer

Yes, in development. The /preview route runs the same screen from an in-memory simulator with no credentials, taking parameters like total=300 and rate=20, so you can watch sorting happen before touching Google Cloud. pnpm test runs a five-sample classifier check and skips itself when there is no Gateway key, and pnpm sort:reset forgets sorting and corrections while keeping fetched messages.

Question 7

What licence is it under, and what is it built with?

Accepted Answer

MIT. The stack is Next.js 16, Auth.js, Drizzle with SQLite, AI SDK 7 for the experimental_evaluate call, Motion, shadcn/ui, and Tailwind 4. Because Jev is reached through the Vercel AI Gateway, the same call path also works with Gateway credits or another evaluation model the Gateway exposes. The project is not affiliated with TypeSafe AI, Vercel, or Google.

Jevmail — fazlerocks

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

Other listings of this project

From the Jev guides

From the Jev AI model community