Features
Email, tasks, projects, calendar, contacts and files, joined up and fast. Below is each area in detail, with the screens as they actually look and the agent’s part in each one spelled out.
Every AI claim on this page is something WI1 does today, not a roadmap item.
As many mailboxes as you like over IMAP and SMTP, or Gmail and Outlook over OAuth. The difference is not the connection — it is what has already happened by the time you look.
assets/screens/email-triage.pngassets/screens/email-verify.pngNew mail is categorised needs reply, FYI or low priority and the label goes on the message, not into a separate view you have to remember to open.
Scored, moved to a Spam folder, and told you why. Your own allow and block rules work per sender or per whole domain, and a quarantine can be cleared one message at a time or all at once.
Reply, task, deal, calendar entry, file it, or leave it alone. One email can yield up to five extra tasks or diary entries, and you accept or dismiss each one separately.
Written from the last fifteen messages, the contact record and your own knowledge base. It lands in the composer: nothing is saved, nothing is sent.
Delegate a thread and the agent works it through, then puts it back on the thread marked awaiting review. You complete it, or send it back.
Switch it on per mailbox and every send is checked against the thread and the attachment text first. A message going to the wrong person stops on a verify page — fix the address, send, or bin it.
A thread links to the contact, the task, the project and the deal, so the history is one record. Archive, delete, mark spam or link a whole selection in one go.
A signature per mailbox, inserted by the composer above the quoted reply. Schedule a send, cancel a scheduled one, or open the raw HTML when a message renders oddly.
When an email contains details that belong somewhere safer, one button reads them out into a pre-filled vault entry for you to check and save.
Seven separate passes work in here: triage, spam scoring, suggested actions, reply drafting, delegated threads, the wrong-recipient check and vault extraction. Each has its own switch.
Monday, 08:40
Half an hour of sorting before any actual work starts — and the one that mattered was four screens down.
It is already sorted. The junk is quarantined with a reason, and the handful that need you are the ones left.
How: Two worker passes run over new inbound every ten minutes — one categorises, one scores for spam. Both are switches you can turn off.
The near miss
Autocomplete picks the wrong address and you find out when a stranger replies asking who you are.
The send stops and tells you exactly what looked wrong about it. Correct the address and send, or bin the whole thing.
How: All six human send paths go through one guard that reads the body and the attachment text against the thread. Scheduled sends and out-of-office replies are exempt by design.
Tasks & AI proposals
This is the part no other CRM has. Write the task the way you would write it for a person. Before you have finished reading it back, it has been assessed.
assets/screens/task-ai.pngassets/screens/kanban.pngEvery new task — however it was created, by a person, a form or an agent — is assessed for how much of that particular job can be taken off you.
What it intends to do, in plain words, before anything starts. Approve it, decline it, or answer the question it came back with.
If your own written rules say a job of this kind needs a deadline or a detail, it asks for that rather than inventing one.
Approved work is logged on the task as an activity entry, so whoever picks the job up next can read exactly what was done and what was not.
Assessment is a proper queue with a lease, swept every minute, so an interrupted one is retried rather than quietly forgotten.
Subtask descriptions are written for you, and a project can propose its own task list — in order, with the obvious dependencies — for you to edit.
Drag between columns and filter by project, holder, company, contact or due date. Invert any filter to see what it excludes, and save the views you return to.
Deals sit beside the tasks and the mail, so “the quote that went quiet” is something the CRM knows about rather than something you have to remember.
Bulk operations take a restore point first, so a mass edit that went the wrong way can be put back.
Working in here: assessment on creation, the task agent that carries out approved work, written subtask descriptions, and project planning.
Thursday, 11:40
You block out the afternoon, then discover two hours of it was copying things between two windows.
“I can do 70% of this” — and a proposal saying which 70%. You keep the half hour that actually needed you.
How: A task assessor reads the task text and anything you have already told it, then writes the percentage, the proposal and any question onto the task itself.
The project nobody scoped
A quiet agreement that someone will break it down next week. Next week, it is still one line.
It proposes the task list and you edit it — a different and much shorter job than starting from nothing.
How: Project planning proposes tasks on the project’s own plan page for review. You accept, change or bin them; nothing is created behind your back.
AI chat
Everyone gets their own agent with its own memory. It does not just answer questions — it has 47 tools and it uses them.
assets/screens/chat.png47 tools in all. A sample of what they reach:
Its memory is yours. It builds up as plain Markdown files on your server that you can open, edit or delete — not a black box you have to take on trust.
The agent inherits the permissions of the person it belongs to. A manager sets what each one may see and do, and it cannot reach past its own user or into another company.
Creates the task, moves the deal, books the slot, drafts the reply, writes the note. The tools are the point.
Email bodies, file contents, tasks, deals, the calendar and your knowledge base — from one question, instead of six searches.
It can search and read your own earlier chats, task chats and machine messages. Only yours: other people’s are out of reach by design.
Documents are text-indexed on upload. When a PDF turns out to be a scan with no text layer at all, the pages themselves are read for that one answer.
Work better done on a machine of yours goes to the agent running there, with the instruction rewritten in full so nothing is lost in the handover.
Dictate instead of typing when you are on site or in the van, and get on with the driving.
The chat agent is the same engine the rest of the CRM uses, with the same permission boundary. What it cannot do, you cannot do.
Finding things
Four searches across email, the shared drive and someone else’s laptop. It was in a thread from March.
You ask for it in plain words and it searches mail bodies, files, tasks and the calendar together, then hands you the one you meant.
How: One agent, 47 tools, scoped to exactly what you personally are allowed to see.
Paperwork
Somebody skims it, or retypes the bits that matter, and the answer you needed was on page 31.
Ask what the termination clause says and it quotes you the paragraph.
How: Scanned pages are read for that question when there is no text layer to index — capped at twenty pages, and the transcript is not kept afterwards.
Projects, calendar, files & contacts
The AI is the reason to move. This is the reason to stay: the everyday parts are fast, uncluttered and work properly on a phone.
assets/screens/projects.pngassets/screens/calendar.pngassets/screens/contacts.pngProjects, their detail and their own planning page; subtasks, deadlines and the dependencies you can actually see.
Month and week views, recurring events, reminders, and company-wide entries that either block everyone’s availability or simply sit alongside it.
Events sync both ways, and the mirrored ones stay visibly mirrored, so you can always see which calendar a thing came from.
Upload them, or keep them from email. Documents are text-indexed, so a search finds the sentence rather than just the filename.
A share link for getting a file to someone outside the company, or across to another machine, that expires on its own.
The intel panel summarises your own history with a person. When it gets something wrong you correct it, and the correction is taught to every later pass.
Two records for the same person get joined into one, rather than argued over by two people in two tabs.
The things you refer to, the links worth keeping, and the credentials that should never have been in a spreadsheet.
Laid out mobile-first, not a desktop grid squeezed down until it technically fits.
Quieter here, by design: contact summaries and task descriptions are written for you. The calendar and the files are plumbing, and plumbing should be boring.
Contacts
Five minutes scrolling your own sent folder, plus a guilty look at their website in the car park.
Open the contact: what you have discussed is already summarised, and anything you have corrected before stays corrected.
How: Contact summaries are built from your own email history and refreshed in the background. Corrections are stored separately and fed into every later pass.
Files
A 30 MB attachment that bounces, or a link to a folder nobody can find again next year.
A share link that works immediately and expires on its own.
How: Files and email attachments live in the same place; a link is generated per file and times out without anyone having to remember to revoke it.
Also in the box
None of these need a sales pitch. They are here because a CRM without them is a demo.
People and the businesses they belong to, with duplicate merge and a summary panel per contact.
Deals with stages, sitting beside the mail and the tasks they came from.
The things your team keeps re-explaining, written once — and readable by your agents.
The links that live in six people’s browsers, in one place instead.
Logins and keys with proper access control, rather than a spreadsheet called passwords.
Attachments and uploads together, with expiring links for anyone outside.
Owner, manager and staff, deciding what each person — and so each agent — can reach.
Dictate a task, a note or a reply instead of typing it.
A proper one, not an inverted screenshot.
One unread alert per thing, replaced rather than repeated.
Multi-tenant by design, self-hosted by choice: your database, your backups.
A restore point before any bulk operation, so a wrong move is reversible.
The controls
An agent you cannot supervise is a liability, not a teammate. This is the full list of gates, and underneath it the three places work happens without one.
And the three exceptions, because you would find them anyway. Mail the AI has classified as junk can be filed or deleted without a click — but only if you switch that on. An agent on one of your own machines, once you have enabled task creation, creates real tasks without review (rate-limited by the hour and by the day). And a connection you have deliberately set to “post” rather than “propose” publishes straight away. Everything else proposes first.
How it works
Add your mailboxes — IMAP and SMTP, or Gmail and Outlook over OAuth — plus your calendar and contacts. Invite the team and set what each person can reach.
Write it the way you would write it for a person. It is assessed straight away: how much of it can be taken on, and what the agent needs from you first.
Give the nod and the work starts. Big jobs get split into subtasks — some for the agent, some for you — and what it did is written on the task.
Corrections stick. What your agent learns about your work is kept as readable files on your own server that you can open, edit or delete.
Machine activity
2 running
Hermes-01 · Rendering export
● running 4m
OpenClaw-A · Scraping feed
● running 1m
Hermes-01 · Nightly backup
done · 38s · $0.02
Hermes & OpenClaw
Send tasks to your Hermes-agent and OpenClaw PCs — by hand, or let them pick work up themselves. WI1 becomes your control room instead of a Discord channel.
Join the waitlist to get early access, or talk to us about connecting your Hermes and OpenClaw machines.