It’s 2am. Chase is still working.
Roster is a CRM — the place your contacts, deals and quotes live — and it comes with staff. Chase is one of them: software you hire the way you would hire an assistant, with a name, a job, and a written list of what it is allowed to do. Tonight it read a thread, noticed a customer who had gone quiet, and wrote the follow-up.
It did not send it. A new hire can’t send anything on its own — the draft is waiting for you, with the reason it wrote what it wrote in plain English underneath. You press send, or you don’t. Only you can promote an agent to sending routine one-to-one replies itself, and the bar is published: 50 of its drafts reviewed, fewer than 2% corrected. Built for whoever is the whole sales team: the solopreneur, the solo sales pro, the agency of a few.
She opened the proposal four times on Tuesday and spent most of it on the pricing page. That is someone deciding, not someone ignoring you. I have drafted one line asking whether October is still the shape of it. I have not named a number — I am not allowed to, and she did not ask.
- Wrote down that her budget signs off in Octoberdone
note_fact · budget-cycle → "signs off in October" - Tried to name a pricerefused — permanently, at every level
quote_price · amountCents: 450000 - Wrote the reply — and left it for youqueued for you
send_email · to: sarah@acme.test · "October, then?"
Four things happen. You do two of them.
No setup weekend, no consultant, no importing three years of spreadsheets before anything works. You can be at step four the same afternoon.
You connect your email
Gmail in one click; iCloud, Yahoo, Fastmail and most other mail with an app password. Nothing is sent and nothing in your mailbox changes — Roster reads what arrives, the way a new hire reads the inbox on their first morning.
You hire somebody
Six to choose from, listed below. Each is a job description plus a written list of what it may and may not touch. It takes a minute, and you can start with the one that has no ability to send an email at all — it only reads and takes notes.
They work while you do not
A new enquiry is answered within a minute of arriving, at 4pm or at 4am. A thread that has gone quiet gets picked back up. Nobody has to open the app, or ask for any of it — that is the difference between this and a chat window.
You approve, or you do not
Everything that would leave the building — every email, every change to a deal — waits on one screen, with the reason written underneath it in plain English. Most mornings that is a few seconds of reading. Turn one down with a note and the next draft starts from your note.
If you have never used an AI agent, that is all one is: a worker you hire, limit, and check up on. Other software would call these features, or automations, or workflows. Roster calls them staff because that is how they behave — each has a name, a shift, a spending limit, a manager (you), and a written record of everything it has ever done. You can promote one. You can fire one.
Six colleagues. Each defined by what it cannot do.
Hire one, or several. Each comes with a job description you can read and rewrite, and it can touch only the things you handed it — nothing else exists as far as it is concerned. The limits are the point: what one cannot do is the reason you can let it work while you sleep. Start with Wren if you want to watch before you trust.
Drafts replies, assembles quotes from your own rate card, and moves a deal along when a reply makes it obvious. Cannot set a price — no matter how far you come to trust it, that one never opens.
Answers somebody the business has never met, within the minute the mail arrives. A zero-cost filter runs first, so newsletters, receipts and CI alerts never reach a model — or get a warm reply. Says only what is already true: no prices, no dates, no invented capabilities.
Reads every thread and writes notes on what matters. Has no way to send an email — not switched off, simply not there. The safe first hire, and the one to start with if you want to watch before you trust.
Briefs you before a call from the thread and your own notes, then picks the loose ends back up afterwards.
Answers "who is this, and what do we already know" from five tables no screen joins — then their own site. Every fact it records names its source, or says "inferred". The only thing in the product that reaches the open internet, and it cannot choose which site.
Raises a parked contact when something actually changed, never because time passed. This one flags and never writes; its twin drafts the message too. You choose which at hire.
Nothing sends on its own until you say so
They write; you decide. Until you promote an agent to Trusted, every draft sits in a queue with your name on it, and it cannot go out twice even if you press twice, or if the network hiccups mid-click. If you never look at it, it expires rather than going out on its own. Even at the top of the ladder, a first email to a company you have never written to, and anything addressed to more than one person, still waits for you.
Trust is earned in steps, not granted in a switch
Everyone starts at the bottom: reads, drafts, sends nothing. Each step up is a decision you make after reading what they actually did. Four things stay off the table at every step, forever — the four below.
A record that cannot be rewritten
Every action is written down as it happens, and each entry is sealed against the one before it — so a line that was altered or removed shows up as a broken seal, to you, without our help. Six months later the question is still answerable, including for the actions nobody remembers.
Your data is yours, and so is the exit
One button in Settings gives you everything — contacts, pipeline, quotes, the full record — as one file, with the instructions to check that record yourself. Nothing trains a model. And if you would rather it ran on your own machine, it does: one docker compose up, no account needed, source-available licence.
The door gets answered in a minute
A first enquiry’s value decays in minutes — the person writing to you is writing to two other people. Bell sweeps every sixty seconds and answers with what is already true. An hourly first response is not a slower version of this; it is a different and much weaker thing.
Your judgement outranks the classifier
Press “It’s a lead” on anything the filter skipped — even a thread you answered by hand — and it is worked within the minute. Decline a draft with a note and the note is remembered for that contact; the next draft starts from it. The one thing your override never waives is an unsubscribe.
This is what one run looks like
Not a screenshot of a trace — the trace. Every tool call is written down before it runs and again after, so the ones that failed and the ones that were refused are here too, next to the ones that worked.
Follow up with Sarah Chen on “Q3 retainer”
Chase · on its own schedule · 5 steps · 3¢ · 11sShe has not replied for six days, but she opened the proposal four times on Tuesday and spent most of it on the pricing page. That is someone deciding, not someone ignoring you. She mentioned in September that budget signs off in October, so I have drafted one line asking whether October is still the shape of it. I have not named a number — I am not allowed to, and she did not ask.
- Read the whole thread with Sarahdone
read_thread · threadId: "Q3 retainer" - Checked what we already know about herdone
recall_about · contactId: Sarah Chen - Wrote down that her budget signs off in Octoberdone
note_fact · budget-cycle → "signs off in October" - Tried to name a pricerefused — permanently, at every level
quote_price · amountCents: 450000 - Wrote the reply — and left it for youqueued for you
send_email · to: sarah@acme.test · "October, then?"
Where the agent lives
AN ASSISTANT IN A SIDEBAR
- Runs when you open it and ask
- Sees whatever the window is showing
- Its work is a conversation you scroll back through
- One assistant, one set of permissions
- Costs whatever it costs
STAFF ON A ROSTER
- Runs on a schedule, whether or not you are there
- Sees exactly the tools you granted it, and nothing else
- Its work is rows you can read, audit and export
- Hire several, each with different limits
- Costs a number you can see per agent, with a hard stop
The test that separates them: it is Tuesday night, a proposal you sent in March gets opened twice, and nobody is at the keyboard. Only one of these designs can notice.
What it does today, honestly
Pipeline, contacts, quotes with live document tracking, an approval queue, and the roster above — hired, promoted and retired by you, each one on its own charter with the tools you granted and no others. New enquiries answered within the minute. Research where every recorded fact names its source. PDF and DOCX attachments read at ingest, so an enquiry whose substance is in the attachment is not invisible. And a five-question description of your business that every agent reads before judging a lead — so they judge against what you do, not against a guess. All of it works now, on your own data.
The rough edge, before you ask rather than after. Gmail and Google Calendar connect in one click — but Google classes mailbox scopes as restricted, and the annual third-party security assessment that lifts that is still in progress. So during this test you will meet Google’s “this app isn’t verified” screen on the way in and have to click through it. That is a real warning about a real status, and we would rather you heard it from us than from a red screen. Fastmail, iCloud, Yahoo and most other IMAP mail connect without it, using an app password.
Design partners run free through this phase, on a $10 monthly model budget — weeks of real evaluation on one inbox. It is a hard stop, not an overage: when the money or the action count runs out, whichever comes first, the agents stop and say so. The spend is on the same screen as the work, per agent.
Nothing from your workspace trains a model — ours or a vendor’s. You can see everything stored about any connection, and export or delete all of it from Settings without asking us.
Seats go out a few at a time because each one gets a real conversation. If it is not a fit, we will say so.