Open source · Inbox triage

Wake up to an inbox that has already drafted the replies.

An AI agent that reads every thread overnight, separates what genuinely needs you from what doesn’t, and leaves a reply waiting in your drafts — written in your voice. It never sends. You press send. The whole thing is open source, so the six decisions below are the ones that actually make it work.

See a night’s run Read the source
01 — how it is built

Six decisions, in the order they matter.

  1. 01

    Grant drafts, never send

    Connect it to Gmail with the drafts scope and leave send out. Sending is not forbidden by a prompt — it is impossible at the API. That one decision is what makes the rest safe to trust.

  2. 02

    Split action from category

    Two separate axes. Action is a closed set — reply now, reply later, no reply — and it alone decides whether a draft gets written. Category (client, prospect, supplier) is open, and is report metadata only.

  3. 03

    Learn the voice from sent mail

    Do not describe your tone — extract it. Point it at a couple of hundred of your own sent replies and let it build the profile: greeting, sign-off, sentence length, the phrases you actually use.

  4. 04

    Flag, don’t guess

    Anything that would commit you — a price, a date, a scope, a yes — gets flagged for you rather than drafted. The agent is deliberately good at knowing where its judgement ends.

  5. 05

    Grade it with fixtures

    Twenty test emails with the answer you expect, run on every prompt change. It turns “seems better” into a number — the difference between tuning it and guessing.

  6. 06

    Close the loop on edits

    When you edit a draft before sending, diff what was written against what went out and fold that back into the voice profile. The drafts get closer every week without you writing a single rule.

02 — see it running

One night, 66 threads, nothing sent.

The demo below is a real run rendered from the same report shape the live system validates before it writes a digest — replies drafted, commercial questions flagged rather than guessed at, and the noise reported as a count instead of a list.

Open the demo workspace
Recorded · Instagram Your inbox, handled before you open it It drafts. It flags. You decide.
Watch this on Instagram
03 — from here

Two ways to go from here.

Build it yourself

Everything above is on GitHub

The full working setup — the triage skill, the prompts, the voice-profile builder and the fixture suite. Clone it, point it at your own inbox, and have your AI editor walk you through the setup. MIT licensed.

github.com/cuprixx/emailtriage

Or have it built

I’ll set it up on your inbox

A free 45-minute call on your actual inbox: whether triage is even the right first automation for you, what it would take, and what it would cost. You leave with the plan either way.

Get in touch