Native macOS · v0.5.0 · Apple silicon

Eight emails matter
today. It finds them.

Inbox Sanity reads your new mail, flags the few things that actually need you, and handles the rest the way you would — on rules you wrote in plain English, with a logged reason for every decision and an undo for anything reversible.

macOS 26+ · Gmail or any IMAP/SMTP account

The Overview screen: '8 emails need you', counters for inbox, unread, needs attention, and approvals, and a list of flagged emails, each showing why it was flagged and an answer found in earlier mail.

Overview. Start each day with a short list of the emails that need you, each with the reason it was flagged and answers found in your own mail.

The Inbox: each email lists how many rules checked it; the open email shows an AI Summary, Related correspondence, and past correspondence with the sender.

Inbox. Every email shows which rules checked it, and opens with an AI summary, related emails about the same matter, and your past emails with the sender.

A drafted reply to Maya Chen that gives the PO number from an earlier email, ready to edit and send, with the original message shown alongside.

Drafted reply. Inbox Sanity drafts replies from facts in your earlier emails, in your own writing style, and you edit and send them.

Newsletters: a TLDR AI issue split into a list of articles with three sponsors hidden, and one article open in a clean reader view.

Newsletters. Newsletters skip your inbox and arrive as a list of articles, with sponsors hidden, that you can read in a clean reader view.

Approvals: an email waiting for approval to unsubscribe and move to Trash, with a 'Why this happened' list of each check made and the model's confidence.

Approvals. Before a rule acts, it can wait for your approval and show each check it made and how sure the model was.

Action rules: four rules such as Phishing and scams and Cold sales outreach, each a plain-language question with its actions, a mode of Automatic or Ask first, and a seven-day count.

Rules. Write rules as plain-language questions, and choose for each one what it does and whether it acts on its own or asks you first.

Activity for the last 7 days: 193 emails handled with a daily bar chart, a breakdown by rule, and a list of actions taken such as archived, replied, and unsubscribed.

Activity. See how many emails Inbox Sanity handled, which rules did the work, and what it did to each one.

The whole idea, in three beats

It shows its work.

Most tools either sort mail into folders you still have to read, or hand an AI your mailbox and hope. Inbox Sanity does neither — it runs on your rules, and you can read exactly why it did what it did.

01

Your inbox has eight things that matter.

Out of 28 new messages since yesterday, only eight genuinely need you today. The rest is noise wearing a subject line.

02

It finds those and deals with the rest.

Flagged mail is surfaced; everything else is handled the way you would — archived, labeled, unsubscribed, rolled up.

03

And you can take it back.

Every decision is logged with its reason and the model's confidence. Disagree? Change the rule. Undo what's reversible.

Rules you can read

A question, plus
what to do about it.

A rule is a yes/no question in plain English. Around it sit deterministic conditions that cost nothing to check — the model only answers the hard part.

  • Action rules compete. First match wins, so a message gets one outcome — not five.
  • Attention rules don't. Each is checked on its own, to catch the thing you'd hate to miss.
  • Flagged mail is never touched. Anything that needs you is held for approval, full stop.

Say “your rules,” not “AI.” The model answers the question — you wrote the question.

The Inbox Sanity rule editor for a 'Cold sales outreach' rule: a plain-English question 'Is this email unsolicited sales email?', the emails it looks at, and a confidence threshold slider set to 85%.
Lead with the brakes

Nothing happens without you first.

Every competitor claims the automation. Few can claim the brakes — the approval queue, the undo, the reasons, the rate limits. Inbox Sanity starts by asking permission for everything and earns automation one rule at a time.

Approval first, automation later

New rules — and anything the Assistant proposes — start in approval mode. Matches queue up in Approvals with the email, the rule, the confidence, and the plan, until you confirm.

Undo on anything reversible

Take it back. A sent reply and a delivered unsubscribe can't be undone — and the app is honest about which is which rather than pretending otherwise.

Rate limits that fail safe

Replies and destructive actions are capped. If a limit trips, every rule falls back to waiting for approval instead of charging ahead at speed.

An “always skip” list

Before any rule runs: Contacts, threads you've replied to, starred mail, your Mac's address book, and anything you safelist. Add “Stripe” and it recognizes mail sent on Stripe's behalf, not just stripe.com.

Trash, never delete

Rules, the Assistant, and MCP apps can only move mail to Trash. Permanent delete is one button on one email, with a confirmation.

The Assistant

Describe it.
It writes the rule.

A chat pane that works on your actual mailbox — reads attachments and PDFs, searches your mail, and writes rules for you. Say what you want in a sentence; it fills in the search, the question, the threshold, and the actions. You review it and save it. Saved rules go into approval mode, never straight to automatic.

It listens and talks

On-device speech with Parakeet & Kokoro on the Neural Engine — or OpenAI's Realtime API if you'd rather. Nothing it drafts leaves your Mac without a confirmation.

Cleanup pass

“Retailer promos, but keep order confirmations.” It writes a filter, shows the matches, and the whole bulk pass can be undone. Never touches Spam or Trash.

Your writing voice

It reads your sent mail and builds a short style guide — greeting, sign-off, length, formality. Quotes and forwards stripped first, so it learns your words.

Your MCP servers

Connect your own HTTP MCP servers and the Assistant uses their tools — each with an ask / allow / deny policy, defaulted from read-only vs. destructive.

Assistanton-device
Stop letting newsletters pile up — but keep anything from my team.
Here's a rule. It runs on messages with a List-Unsubscribe header and skips your Contacts:
PROPOSED RULE
searchhas:list-unsubscribe
question“Is this a bulk newsletter?”
confidence≥ 0.80
thenRoll up · Mark read · Archive
Starts in approval mode — nothing runs until you confirm.
Choose your model

On-device first. Hosted when you want it.

Classification is pluggable, and the on-device options are first-class. Any rule can name its own model — the expensive one handles judgment calls, the local one handles volume.

On your Mac

Runs fully offline
Apple Intelligence — defaultOllamaLM StudioTev1 on MLXOsaurus

Hosted

You choose · bring your own key
AnthropicOpenAITypeSafe JevCloudflare ClefOllama on another machine

A rule whose model isn't set up is skipped — it never quietly falls back to a different one, because a confidence threshold tuned for one model means nothing on another.

Also in v0.5.0

Every Mac. Every email. Your other apps.

Run it across your Macs, keep a year of mail searchable offline, and let the AI apps you already use work with your inbox — on the same terms as everything else.

Use it on every Mac

One Mac checks mail and runs your rules, so two Macs never act on the same email; if it goes quiet, another takes over. Approvals and Undo work from any Mac.

Offline mail and search

The last year of your mail, stored and searchable on your Mac, even offline. Encrypted mail is never stored.

Inbox Sanity as an MCP server

Let Claude, Cursor, or any MCP app work with your mail. Read-only by default; with changes turned on, each one waits for your click. Never sends mail.

Privacy & security

Your mail stays on your Mac.

No servers of ours sit in the middle. The app talks to your mail provider, the models you choose, and only the other services you turn on.

local only

Everything lives in an encrypted database on your Mac (SQLCipher, with a key in this Mac's Keychain that never leaves it). Nothing goes to servers of ours, because there aren't any.

icloud sync

Optional and off by default. Turn it on per Mac to sync rules, settings, and activity through your private iCloud database, end-to-end encrypted. Email bodies and OpenPGP secret keys each have their own switch, also off by default.

offline capable

Every model can run on your Mac: Apple Intelligence, MLX, Ollama, and on-device speech. Nothing has to leave it.

keychain

Passwords, tokens, and API keys live in the macOS Keychain, never in the database. With iCloud sync on, they move only through iCloud Keychain. The app is sandboxed, signed, and notarized.

openpgp

Encrypted mail is decrypted in memory and never written to disk, unless you let a Download attachments rule save decrypted attachments. Off-device models don't see decrypted text unless you explicitly turn that on.

safe compose

The composer won't send a reply to an encrypted email in the clear unless you turned encryption off yourself — and refuses to encrypt when a key is missing.

key discovery

Secret keys go in the Keychain; your contacts' public keys are discovered via WKD and Autocrypt, so encrypted mail just works.

your data, your call

Every feature that sends data off your Mac asks first, and you can withdraw that permission. Delete all local data or all iCloud data with one button; removing an account revokes Google's access too.

Menu bar extraSiri & ShortcutsMorning briefingFollow-up watchDaily Summary & roll-upsNewslettersRelated correspondenceSummary pane
Read the full privacy policy →
Early access

Give your inbox back to you.

Join the early access list. We'll let you know the moment Inbox Sanity is ready for your Mac — no noise, which would rather defeat the point.

Thanks — you're on the list. Watch your inbox (it'll be one of the eight that matter).

macOS 26+ · Apple silicon · Gmail or any IMAP/SMTP account · Privacy Policy