The homepage covers the three things nobody else can give you. This page is the rest — the full toolkit of a working communications agent across email, calendar, and Slack.
Four verdicts — important, learn from, not important, archive — applied by rules it learned from you, not filters you had to write.
Correct one call in plain words and it becomes a durable rule — exceptions included.
It remembers how you treat each sender and applies that knowledge only when that sender shows up.
Prospecting mail gets flagged and remembered. One click purges a sender for good.
Newsletters you never read get unsubscribed, not just archived.
Newsletters worth keeping are harvested into knowledge, then archived — read without reading. And harvesting isn’t limited to newsletters: any email carrying a durable insight — a client sharing what works, an operator changing policy — gets distilled into the knowledge base while the mail is handled normally.
Filed mail stays visible while its task is open, then archives itself when the work is done.
Email is work in disguise. Letterknife makes the work explicit.
Every real conversation is tracked as a task with status, history, and a next step.
Each task shows what is currently being asked of you — not just the last message.
Related threads roll up into goals, so pursuing one objective through three people is one item, not three.
It knows when someone replied, when they didn’t, and chases the silence for you.
When the ball is on their side the task parks as waiting — and reopens by itself the moment a reply lands or its follow-up comes due.
“Chase in a week” is one sentence: the task parks, and when the date hits the agent checks the thread and drafts the nudge if they stayed quiet.
Tasks group into your projects — a name, keywords, and optionally a working folder, so code tasks run inside the codebase they belong to.
Things others are driving get an eye icon and stay out of your plan until they need you.
See when something first came up versus when it last moved — nothing quietly rots.
The inbox turned into an operated day.
Calls, replies, approvals and chases assembled into one prioritized list, every day.
The plan shows how your hours split between calls and desk work before you commit to it.
Meetings sync in near-realtime — moved, cancelled, and declined calls resolve themselves.
A 24-hour digest of everything the agent did, so unattended never means unaccountable.
Nothing you did yourself rings the bell — notifications are reserved for what happened while you weren’t looking.
Drafts in your voice, sent only with your consent.
Replies are written from your history and style — ready before you open the thread.
Ask it to write someone and it composes a fresh email in your voice — drafted for review, with a follow-up task on the new thread so the reply comes back tracked.
Nothing sends without an explicit OK. Typed conversation never counts as permission.
It reads image attachments for context, and any attachment opens natively with one click.
Drafts account for the whole conversation and the sender’s profile, not just the last message.
Sending a draft is treated as your move: the task updates itself — ask cleared, follow-up scheduled, status parked — without another click.
One brain across the places work actually happens.
Your login is your identity; soon you’ll connect every inbox you own under one plan — work, personal, side project.
Slack conversations flow into the same tasks and plan as email — one thread of truth per topic.
Every sender gets a profile: open tasks, history, disposition — visible right in the reader.
Sweep an entire backlog of senders in one run once the agent knows your preferences.
The onboarding is the training. No forms, no interview.
Resumable training slides: it judges, you correct, it converges — visibly, in one sitting.
“Usually skip — except my own community” is one correction, captured whole.
A running score of accepted recommendations shows the learning actually happening.
Everything it learns lands in a knowledge base it curates itself — general knowledge, plus facts about the companies and people you deal with, each surfaced only in the right conversation. Every entry in plain language, editable or deletable in the app, mirrored to plain-text files on your disk. No black box.
On a schedule you pick (daily or weekly), a maintenance chain folds in your recent corrections, audits the whole knowledge base for contradictions and stale facts, dedups what it harvested, and lifts durable facts to where they belong. No gardening required.
Most tools have an API about the product. Letterknife exposes the product itself to your AI.
Every capability — reading, tasks, calendar, sending — is a tool your own Claude can operate.
Drop an image or PDF into any task chat and the agent reads it directly.
It budgets your Claude subscription like it budgets your time: visible spend, auto-pause, resume.
The agent, its memory, and its rules live on your machine. Close the lid and nothing of yours is anywhere else.
An assistant that knows your life is a target. Everything Letterknife knows carries a disclosure level, and what unlocks is decided in code from verified sender addresses — never by what an email claims.
Who you are, how you work, reference knowledge it has harvested — everything safe to use in any conversation.
Unlocks for: everyone, always.
Account details at your bank, procedures with your insurer, project state with a client. Family works as a group too.
Unlocks for: mail sent from an address that organization vouches for — double auth: the right person and the right address.
What you know about one specific person — history, preferences, the personal record of an ex-colleague or a friend.
Unlocks for: that person only, from their verified addresses.
“Your broker” writing from a random Gmail address meets an assistant with no account numbers in its context and no tool that will hand them over — the best it can do is draft “I’ll check and get back to you.” Your real broker gets the full picture automatically.
What unlocks is decided by verified sender addresses, never by what an email claims — gated facts are never even injected, and the AI can’t widen its own access. Only an instruction you type yourself overrides.
When someone leaves a company and their work address is shut down, their access to company info ends with it — their personal record still works, the company’s doesn’t. No revocation list to remember.
“I’m from your bank” creates nothing. A new company relationship takes real engagement from you — you actually replied — plus a person who verifiably exists at that company. The impersonator whose “employee” isn’t findable anywhere is caught before anything is created.
The agent can harvest a useful insight from any email — but that write path is restricted in code to adding inert, public-level notes. A malicious email can never use it to alter what the agent knows about you, a company, or a person.
An optional security mode strips the agent down to Letterknife’s own tools — no filesystem, shell, or web — so a malicious email can never reach your files, your repos, or the wider internet.
The full model — and why it's a guarantee rather than a guideline — is in the security docs.
Being honest about what's not here yet:
The feature list is one thing. Watching it learn your mail in twenty minutes is another.