Email, calendar, matters, people and files, gathered into one workspace on your own machine. A team of named agents, each with clear limits, reads, follows up, drafts and files — and only what needs your decision reaches you. Every step traces back to its 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…
Client follow-up wants to run Draft reply — your approval is required.
Today gathers what needs you: approvals waiting on a decision, the day's meetings, replies you owe, matters coming due, and whatever your agents just finished. One read, and the shape of the day is clear.
Not batch clean-up after the fact — every email is processed as it lands. Category, priority, a one-line summary and the next step are written into each message as structured fields, so the list tells you what to open first.
Emails are clues; the work is what you're actually moving forward. A matter keeps the background and goal, action items, stakeholders, threads, meetings and files on one page. A follow-up agent keeps watch — but it only proposes. You decide what's written.
The North America and EU payment gateways move to a new provider in Q4. Northwind wants a joint review before cutover, with a rollback window agreed in advance.
Both regions cut over on schedule, with cutover-week transaction failures no higher than today.
Each item has its own run state and report. When the agent needs an answer from you, it stops and asks instead of spinning.
The follow-up agent wakes on a schedule, a new email or a meeting's end, gathers evidence and suggests updates. You accept or decline each one.
Progress tells the story of the work; the activity log records who changed what. You read a narrative, not a ledger.
Product lead for payments at Northwind. Your threads with her center on the settlement API migration and the EU launch. She asks for a written proposal before deciding, and usually copies legal and finance. For the past two months she has been scheduling the Q4 gateway cutover review.
Contacts builds a profile for everyone you correspond or meet with, merging addresses into people. The calendar layers mailbox events, matter deadlines and agent schedules into one view. Your agents know who you work with, and when you're free.
Email attachments, chat attachments, documents your agents write, and folders you mount from disk form one file tree. Files stay where they are; the index holds only metadata. Download the semantic model once, and search runs entirely on your machine.
Your main agent has its own name and face, alongside members with distinct jobs — the daily report, meeting prep, matter follow-up. They wake on the rules you set and act within the capabilities you grant; every run leaves a complete record.
The report agent looks back over your mail and matters and writes a structured report: what was handled, what still needs you, what moved forward. Daily, weekly and monthly, from one engine.
Summon the main agent from anywhere. It brings the email, matter or file in front of you into the conversation, searches, reads and compares on its own, and cites what it found.
Letting an agent act for you starts with knowing it won't overstep. MailAgent builds its limits into the structure, not into a prompt.
The assistant wants to run Draft reply — your approval is required.
This will be saved to Drafts — nothing is sent (edit before confirming)
Run directly, ask first, or disabled — set per tool. Frequent, reversible actions like drafting, flagging and archiving run directly by default; the rest ask first.
Sending email, running commands, installing skills and changing agents always ask, and that can't be lowered. The only exception for sending is a recipient allowlist you set yourself.
Recipients, subject and changes on an approval card are generated by the server from the real payload, not paraphrased from the model.
Which folders and files an agent may write is decided by the backend, not by the agent. Unattended runs may only append.
Statistics, links and email IDs in a report are filled in deterministically. IDs the model supplies are checked, and invented ones are dropped.
Mail, indexes and files live on your computer. Only the model provider you configure sees the portion it's asked to process.
Chat with it in Feishu, check Today from a phone browser, let it look things up in Notion, Jira or GitHub, or hand your inbox to an agent of your own. Wherever it runs, approvals and permissions still apply.
Connect 16 services over MCP, including Notion, Jira, Confluence, Slack, GitHub, Linear and Figma. Authorize in the browser; credentials are encrypted on your machine.
The same assistant and memory in a Feishu direct message. Actions that need approval arrive as button cards — one tap to allow. Uses a persistent connection; no public address needed.
Host your own web entry point to read mail, Today and reports on the go. Features that depend on local services, such as Library and group chat, stay on the desktop.
Issue scoped access keys and call mail reading, search, reports and calendar over REST or MCP. Write and send actions require explicit confirmation.
A local-first desktop app with public source. Download, connect your mailbox, choose a model — and it starts with the next email.
Get the latest installer from GitHub Releases: .dmg for macOS, .exe for Windows.
Pick a mail backend in the first-run wizard: DavMail, or local Outlook on Windows. Notion is optional, with one-click authorization.
Add your own model provider key, switch on the agents you want, and it starts with the next email.