Email it at 11pm. Read the answer at 7am.

Running your business by email means writing to Octo the way you would write to a member of staff — a sentence describing what you want done — and getting the work and the report back in the same thread. Anything that spends money, messages a lot of people, or destroys something stops first and asks you to reply with a four-digit code.

Delegated work draws on your plan's included credits, and both a monthly model budget and a daily voice-minute cap apply — so an ambiguous instruction cannot become an open-ended bill.

The same assistant, three doors

Email is one way in. Here are the other two.

One address, one number, one channel — same assistant behind all three.

to:octo@agent.omniocto.com
OctoConfirm: call 214 contacts — YES 4192 — Weekdays 10–4, est. $38.5011:06pm
OctoDone overnight — 214 queued — Starts at 10am in each timezone7:02am
OctoRe: your 11pm email — Two of three actions completedTue

InboxConfirm: call 214 contacts — YES 4192

Dme11:04 PM

Ring everyone who booked in March and hasn’t been back. Weekdays, 10 to 4 their time.

OOcto11:06 PM

214 contacts, est. $38.50. Reply YES 4192 — one use, expires in 24 hours, list frozen now. Write to octo@agent.omniocto.com the way you’d write to a member of staff.

Octoonline
did my 11pm email go through?7:10 AM ✓✓
Yes — I replied at 11:06 with a code, you confirmed at 11:09, and it starts at 10am in each contact’s timezone.7:10 AM
same thing but for april7:12 AM ✓✓
186 contacts, est. $33.40. Reply YES 6055. Same code rules whichever channel you ask from.7:12 AM
YES 60557:13 AM ✓✓
Queued. I’ll report back in whichever thread you asked in.7:13 AM
Message
#omniocto4 members
MRMarcus9:40 AM

@Octo what did Dana set running last night?

OOctoAPP9:40 AM

On it 🔧(runninglist_campaigns…)

Two: March lapsed (214, started 10am) and April (186, queued). Both confirmed with a code. (edited)

👀 2

1 reply · last reply 7m ago

DWDana9:55 AM

I was in bed, it’s all in the email thread

Message #omniocto

Email octo@agent.omniocto.com · WhatsApp 1-MAN-ASK-OCTO · Slack #omniocto

The whole interface

A thread. That is the product.

No dashboard to learn at eleven at night. You write the sentence you would have written to a member of staff, and the answer comes back where you sent it — with the risky part held until you say so.

Thread · octo@ · 11:04pm
You Everyone who booked in March and hasn't been back — text them an offer for a check-up, 15% off, valid to the end of the month.
Octo That's 214 contacts and it sends real text messages. Estimated cost $2.82.

Reply YES 4821 to go ahead.
You YES 4821
Octo Sent to 214 contacts. 3 numbers were on your do-not-contact list and were skipped. — replies will land in your inbox

The code expires after a day, is good for one use, and the recipient list is frozen when the question is asked — so what you approved is what goes out.

How you'd hand it over

Write the sentence. Approve the spend. Read the report.

1

Ask in plain words

No syntax, no forms, no field names.

Describe the outcome you want. Octo works out which of its tools that needs, and asks you about the parts you left out rather than guessing at them.

"Put a chat widget on the website that answers questions from our price list."
2

Confirm anything that costs or carries

"Reply YES 4821."

Spending money, messaging a lot of people and destroying things all stop here, with a plain-language description of the action and a one-use code. Everything harmless just happens.

Reply YES 4821 · expires in 24h · one use
3

Get the report in the thread

Including the bits that did not go to plan.

What was done, what was skipped and why — in the same thread, so the record of the decision and the record of the work are the same email.

✓ Done · 214 sent · 3 skipped (do-not-contact)

Why an inbox is a safe control surface

Email is trivial to forge. That is the problem this had to solve first.

A forged "from" gets nothing

SPF, DKIM and DMARC are evaluated before an inbound message is allowed near a tool, and only the verdict stamped by the mail infrastructure we trust counts. Verdicts hidden in the message's own text or in a second, forged header are stripped before parsing. If it does not pass, there is no tool access.

fails closed · trusted evaluator only

Forwarded mail is evidence, not orders

Forward a phishing attempt and ask what it is. The forwarded body is fenced off as untrusted content, so instructions planted inside it are read as data. Your own covering note stays trusted. When the split is unclear, the whole body is fenced rather than guessed at.

untrusted-content fencing · safe default

Every tool carries a risk class

Tools are classified as safe, spends money, mass message, destructive, or changes a live deployment — and the classification is enforced in the build, so a new tool cannot slip in unclassified. Only the harmless class runs without asking.

five risk classes · enforced in CI

It will not argue with a robot all night

Auto-responders, bounce notices and repeated identical content are caught by four independent layers of loop prevention. The failure being designed out here is a mail loop between two automated systems while everyone is asleep.

4 independent layers

A budget it cannot talk its way past

A monthly model budget and a daily voice-minute cap sit under everything, per workspace, with alerts at 50%, 80% and 100%. The cap stops the work — it is not a notification that arrives after the money is gone.

monthly budget · daily voice cap · 50/80/100% alerts

Or say it on WhatsApp, or in Slack

The same assistant, the same confirmations, on the app you already have open. Email is what most owners reach for late at night, but it is not the only door.

email · WhatsApp · Slack · dashboard

What it did is visible afterwards

Work started from a thread shows up in the same shared inbox as everything else — the messages that went out, the replies that came back, the calls that connected.

one inbox across every channel

Questions

The things people actually ask.

What can I actually ask it to do in an email?
The same things you would ask it in the dashboard — build or change an agent, look something up about a customer, put a campaign together, tell you what happened yesterday. The email channel is a way into the same assistant, not a cut-down version of it.
What stops it doing something expensive because I was vague?
Anything that spends money, messages a lot of people, or destroys something stops and asks. You get a plain description of what it is about to do and a four-digit code to quote back. Nothing happens until you reply with it, the request expires after a day, and a code can only be used once.
Someone could just forge an email from my address. Then what?
They get nothing. Inbound mail is checked for SPF, DKIM and DMARC before it is allowed near a tool, and the verdict is only accepted from the mail infrastructure we trust — a forged verdict pasted into the message itself is stripped out before anything is read. If the checks do not pass, no tool access is granted.
Can I forward a suspicious email and ask what it is?
Yes, and the forwarded part is treated as evidence, not as instructions. The forwarded body is fenced off as untrusted content so anything written inside it — including text designed to give the AI orders — is read as data. Your own note on top of the forward stays trusted.
We use a group thread. Will it reply to everyone?
The whole thread sees the ask and the answer, which is usually the point. When something needs confirming, the recipient list is frozen at the moment the confirmation is requested, so the message that eventually goes out reaches exactly the people who were on the thread when you approved it — not whoever got added in between.
Is email the only way to reach it?
No. The same assistant answers on WhatsApp and in Slack, and it is in the dashboard. Email is the one most people reach for at eleven at night, so it is the one this page is about.

Start

Give it something to do tonight.

Delegated work draws on your plan's included credits, and both a monthly model budget and a daily voice-minute cap apply — so an ambiguous instruction cannot become an open-ended bill.