Montrago is a live mental model of your email habits. It fetches new Gmail messages, categorizes and consolidates them, then predicts the exact action you'd take on each one (archive, reply, snooze, star, label) along with a paragraph of inner monologue explaining why. You either accept the prediction or mind-dump your own thoughts; either way, the mental model sharpens.
Live at montrago.vercel.app.
The Problem
Inbox triage is the same handful of judgment calls repeated thousands of times: what matters, what can wait, what gets archived, what deserves a reply. Generic AI assistants don't know your specific preferences, like who you respect, what projects you're in, or which newsletters you actually read. They give generic answers to deeply personal questions.
The Solution
Montrago builds a markdown document (your mental model) that grows with every triage decision. It captures your core identity, per-category rules, strong opinions about specific senders, heuristics, exceptions, and recent concrete examples. Claude reads it before every prediction. After about twenty triages, predictions unlock: a prose inner monologue, a one-line summary, and a concrete Gmail action. Accept and commit, or override with your own thoughts. Both paths feed the model.
Key Features
- Gmail integration via OAuth: archive, trash, star, label, draft, snooze, flag, or keep, all committed in one click
- Automatic categorization and consolidation (seven job postings become one triage unit)
- Predictions powered by Claude, grounded in a living mental-model document you can view and edit directly
- Every mental-model update is snapshotted so you can roll back
- Accuracy stats track how often your predictions match your actual decisions
- Single-user deployment (set
ALLOWED_EMAIL). Your inbox data never leaves your database
Stack
Next.js 15 + React 19 + TypeScript + Prisma + SQLite. Anthropic Claude for categorization, consolidation, predictions, and mental-model updates. Google OAuth + Gmail API for inbox access. HMAC-signed cookie sessions.