An agent that messages you first · from $600 a month

7:04 am.
Before your
laptop opens.

Every order stuck, every payment failed, every refund overdue — and what it already did about it. One message, then quiet until something needs you. Claude Code, kept alive on a server, inside the chat app already on your phone.

40 schedules running right nowModel usage includedRead-only for the first 14 days
Proactivity

It runs without you

Schedules written in a sentence, not in cron. The brief arrives because the night ended, not because you opened an app — and anything that touches money waits for your yes.

schedules
23:00Send the day's orders, paid and cash on delivery
*/30Abandoned-cart ladder, three touches
09:00Confirm unconfirmed cash-on-delivery orders
08:00Dispatch notices for anything shipped overnight
03:00Nightly curator — fold the day into memory
Daemon23:00 · online
Overnight — Tue 9 Sep

31 orders · $4,218
Shipped 26 · Held 3 · Failed 2

Needs you:
  #4471 refund requested, $86 — yes / no?
  #4488 no scan since Sat

Already done: 2 failed payments retried (1 recovered), 4 Gorgias replies drafted, inventory synced.
7:04 AM
Fig. 1 — the nightly digest arriving unprompted, then a schedule created from a sentence
What it can do

A full Claude Code session, with a body

The same agent you would run in a terminal — shell, files, browser, subagents — kept somewhere that never sleeps.

#1  Perceives

Hears
everything

Send a voice note from the warehouse floor and get a real answer. Transcription runs on our box; replies can come back as audio.

#2  Perceives

Reads
the room

Photos and PDFs are read natively. Send a picture of a shelf and ask it to update the stock count.

#3  Acts

Focused
automation

Written in a sentence: every weekday at 7, list every order with no scan since Friday. It becomes a job.

#4  Acts

Browses
the web

Drives a real browser when there is no API — a carrier portal, an ads dashboard, a supplier login nobody ever integrated.

#5  Persists

Persistent
memory

Conversations, facts and lessons survive restarts, upgrades, and a move to another machine.

#6  Persists

Learns
your store

A task done twice is done better the third time, because it keeps what it learned. On Ops Team, we build the new ones for you.

Before you hand it a store

Trust is earned, so it starts read-only

For the first 14 days it changes nothing. It tells you, every morning, what it would have done. Then you turn it on — and money still waits for your tap.

What it can do
What it can never do
Read orders, shipments, tickets and reports
Your Shopify password — it gets its own staff account, scopes you pick, revoked in one click
Draft replies, open carrier traces, sync inventory
Send a customer anything without your tap
Retry a failed payment, flag a refund request
Issue a refund, change a price or move money without your tap
Run the jobs you scheduled, on our machine
Install software, run as root, or touch its own source on a hosted plan
14days read-only, by default
$8a day model ceiling on Operator, enforced in the software
1tap to approve anything touching money

Every action is logged with a timestamp. File changes are snapshotted before they happen and can be rolled back. If the agent issues a refund you did not approve, we reimburse it, up to three months of your subscription. The source is public under the Elastic License 2.0, so you can read exactly what it will and will not do before you give it anything.

What you get

Your own instance

One Linux account, one daemon, one database, its own scoped credentials. We provision it, keep it patched and keep it running, and we pay for the model. You talk to it.

health
$ daemon health

• Telegram: connected
• WhatsApp: up + linked
• Model: claude-opus-5 (effort=max)
• Schedules: 40 · Workflows: 3
• Backups: nightly, 14 kept
# real output from the instance
# this page was built from
40schedules running
$600a month, one bill
14nights of backups kept

Your instance is yours: its own account, its own database, its own credentials, and no other customer can read it. Cancel and we hand you the data and delete ours.

Pricing

Everyone else sells you a builder

This one shows up already knowing your store, and it messages you first. One bill: model usage is included, usage is counted in jobs rather than tokens, and there is no overage line — a heavy month is a conversation, never a surprise.

Operator

Start here

One store. The 7am brief, every morning, before you open the laptop.

  • Up to 50 jobs a day, unlimited chat
  • Model usage included — one bill
  • Slack, WhatsApp or Telegram, 3 seats
  • Browser automation where there is no API
  • Read-only for 14 days, then approve-first on money
$600
/mo
or $6,000/year
Start on Operator

Ops Team

One store, more of it delegated, and someone building skills for you.

  • Up to 150 jobs a day, unlimited chat
  • The strongest model at high effort
  • 8 seats, shell jobs, HTTP triggers
  • Custom skills built for your store
  • Priority response, same day, named contact
$1,200
/mo
or $12,000/year
Start on Ops Team

Multi-Store

A portfolio, or an operator running several brands from one phone.

  • Dedicated server, 4 vCPU / 8 GB
  • Up to 500 jobs a day, 20 seats
  • Any model, maximum effort
  • One instance per store, one chat to you
from $2,400
/mo
three or more stores
Email me
Add-ons
NamePriceWhat it does
Extra seat$25/moOne more person on the allowlist.
Second store$400/moAnother store on Operator or Ops Team.
Dedicated server$150/moYour own box instead of a shared host.
GPU voice$200/moCloned voice replies and large-model transcription.
Before you buy

The questions with awkward answers

What does it actually cost to run?

The price on the page. Model usage is included and we provision the key; you never open an account anywhere else and never see a token. Each plan is described in jobs a day, and the daemon enforces a daily spend ceiling underneath that. If a runaway job burns through it, that is our cost, not yours.

How do you compare to a VA, or to Zapier?

A VA costs about $2,200 a month loaded, sleeps, and leaves. Zapier and n8n fire a rule or they do not, and the day something unexpected happens is the day they do the wrong thing quietly. Lindy and Relevance hand you a canvas and a Saturday. This is already built, already running on your store, and it opens the conversation.

If you would run Claude Code yourself: this is the same agent, given somewhere to live. It does not stop when you close the laptop.

Is it sandboxed?

No, and we would rather say so than sell you the word. Three things stand in for it. A permission gate blocks destructive commands, credential writes and private-key reads. Anything that touches money or one of your customers — a refund, a price, a payout, a message — waits for your tap and is denied if you do not answer. And for the first 14 days it is read-only: it tells you what it would have done and changes nothing.

An agent that acts on your behalf can still make mistakes on your behalf. Give it a scoped staff account, never your password.

Which chat apps?

Slack, WhatsApp or Telegram. In Slack it answers DMs and @-mentions in the thread, and an approval is two buttons in your DM. WhatsApp runs on a number we help you set up — never your main business line.

Where does my data live, and who can see it?

In its own Linux account on our server, with its own database and its own credentials. Other customers cannot read it. Anthropic sees the content of each turn, under a key dedicated to your instance. Backups run nightly and fourteen are kept. Cancel and we export it to you and delete ours.

Who is behind this, and who answers at 2am?

One person: Sahil. He builds it, provisions it, and is the one you message when something breaks — and it has been running his own store in production for a year. That is a strength at this size and a limit past it, which is why there are three prices and not a self-serve funnel. Timezones are real: response is same business day on Operator, same day on Ops Team.

The source is public. Why pay?

So you can read every line before you hand it a key — that is the point of publishing it. The licence lets you run it yourself, unsupported, and we will not stop you. What you pay for is the machine that stays up, the model bill, the WhatsApp bridge that gets re-linked when it drops, the nightly backups, and a person who answers.

Can I cancel?

Any time, effective at the end of the period you paid for, with a 7-day money-back window on a first subscription. A failed payment drops you to smaller limits after a grace period rather than stopping the agent. Nothing is deleted.

What is not built yet?

No web dashboard, no single sign-on, no roles, no shared team memory, no SMS or email channel yet. No models other than Claude, because the agent is Claude Code. We list these because a feature grid that quietly implied them would be a lie, and you would find out in week two.