Lohnt sich dein eigenes Mailient?

Build a local AI email triage tool called Inbox Warden in this empty folder.

Stack, no substitutions: TypeScript on Node 20, Fastify for the server, server-rendered HTML with htmx for the UI, better-sqlite3 for storage, imapflow for reading mail, nodemailer for sending, and one OpenAI-compatible chat completions endpoint for the AI. No React, no Next, no Docker, no cloud services, no telemetry, no user accounts. This runs on localhost for exactly one person.

Secrets in .env, read with a tiny loader, and ship a .env.example with: IMAP_HOST, IMAP_PORT, IMAP_USER, IMAP_PASS, SMTP_HOST, SMTP_PORT, LLM_BASE_URL, LLM_API_KEY, LLM_MODEL. Never log credentials or full message bodies.

What is in scope:
- A sync job that connects over IMAP, fetches messages from INBOX since a stored high water mark, and stores id, thread key, from, to, subject, date, plain text body and raw headers in SQLite. Dedupe on Message-ID. Make sync idempotent and resumable.
- Strip HTML to text and truncate long bodies before sending anything to the model.
- For each new thread: one LLM call returning strict JSON with category (one of urgent, needs-reply, fyi, newsletter, junk), a two sentence summary, and a suggested reply draft when category is needs-reply or urgent. Validate the JSON with zod and retry once on parse failure.
- Cache results keyed by thread and message count so restarts do not re-bill you.
- A web UI at / with a triage queue grouped by category, thread detail view, and an editable draft box with Approve and Send, Skip, and Mark Junk. Nothing is ever sent without an explicit click.
- Sending via SMTP with correct In-Reply-To and References headers so replies thread properly.
- A rules table for cheap deterministic overrides by sender or subject substring, applied before any LLM call.
- npm run sync, npm run dev, and a --dry-run flag that classifies without sending.

Explicitly out of scope: OAuth flows, mobile app, push notifications, multi-account support, calendar, attachments beyond recording filenames, search beyond a SQL LIKE query.

Write a README with setup steps, an honest limitations section that says polling is not push and that message text leaves the machine for the model, and a cost note estimating tokens per hundred emails. Include a fixture-based test for the classifier JSON parser that does not hit the network.