MailAgent reads every email the moment it lands — classifying, prioritizing, summarizing and drafting replies. What matters rises; noise settles. Every claim links back to the source.
Hi all,
I think the discussion is already thorough — here are a few notes from the field for reference: rollout pacing differs by market and adoption isn’t uniform…
Not a batch job after the fact — processed on arrival. Category, priority, a one-line summary and a suggested action, written structurally into every message so the list tells you what to open first.
Each email is tagged with a category (product, decision-needed, FYI…) and a priority. Critical mail surfaces in red; FYI settles to the bottom — no more digging through 674 messages.
Long threads and deep quote chains, compressed into a point you read in two seconds — before you ever open it.
For mail that needs a reply, a context-aware draft is waiting. Tweak a line or send as-is — you stay in control.
"Needs a decision," "awaiting sign-off," "archive" — it tells you not just what an email says, but what to do with it.
A local FTS5 index over 9,000+ emails. Hit ⌘K for millisecond results — and a graceful fallback when KOS is offline.
Every email carries an AI FIELDS panel up top: summary, priority, action, category, reply suggestion — all structured, reviewable, and markable as "reviewed." Not a black box; a transparent assistant.
MailAgent pushes what's critical to the Dynamic Island. A high-priority email lands, gets triaged, and its gist and suggested action float at the top of your screen — no app, no inbox dive.
Jarvis reviews the last 24 hours and curates a structured report: what's been handled, what still needs you, the key facts you must know, and the rest summarized. Weekly and monthly run on the same engine.
Custom AI bridges to the KOS knowledge base over MCP — email sources, an entity graph, historical trajectories. The model decides what to query from context and answers with citations. We don't spoon-feed it; we hand it the tools and let it drive.
The one thing an email assistant must never do is make things up. MailAgent's discipline: counts, links and email IDs are back-filled deterministically by code; the LLM only handles wording and grouping. The blast radius of a mistake is confined to phrasing.
Totals, unread, urgent, processed, pending — all deterministic and consistent with the database.
Any email ID the model emits is validated against the real candidate set; invented ones are dropped.
Every source link is generated from a real page_id — it always lands on the right message.
Mail lives in local SQLite read directly by Electron; the body is never sent out for translation.
The LLM and system dashboards lay the real numbers bare: throughput, tokens, latency, cache hits, sync health, dead-letter queue. Not decorative charts — the actual ledger of 5,904 runs.
The same design system and the same AI judgments, responsive down to the phone. Focused/Other split, priority flags, island pings — just as clean on a narrow screen.
A local-first macOS desktop app, open source. Download, connect your mailbox, and the AI takes over from the next email.
Grab the latest .dmg from Releases and drop it in Applications.
Connect over DavMail / IMAP; mail syncs to local SQLite.
Add your model key