OCTOVO acts on three adjustable levels — always ask, balanced, autopilot — gives a reason for every draft, writes every action to the log and can be stopped completely at any time with one click.
How it works
It gets to act. You keep the wheel.
Suggestions alone save no time. Doing everything alone is frightening. Where OCTOVO sits in between is up to you — with a slider.
Three positions: always ask, balanced, autopilot.
- One consent, once. During setup it asks once for device access. Never a storm of pop-ups after that.
- One reason, always. Above every draft, one sentence says why it reads the way it does.
- One log, gapless. Including failed attempts and the cases where it wasn’t allowed to do something.
The approval
The outbox where nothing leaves on its own
A draft only becomes a message once you have approved it. What is in front of you when that happens is described here — field by field.
- The whole draft, editable in place
- No second window, no “edit” button. You write in the field that’s right there.
- Why this way — as a required field
- One sentence names recipient, matter and trigger.
- 30 seconds of undo
- After sending, the button sits in the corner. The letter hasn’t gone yet.
- “Do it differently” asks exactly once
- It takes your correction — and asks exactly once: “Remember this for future mails to Ms Berger?”
Three levels
From “asks about everything” to “just handles it”
The level applies to everything at first, then per channel, per order, per contact. “Always straight through to Ms Kern, never to new customers” is one sentence — not a setting.
The lever has three notches: always ask, middle way, autopilot.
-
Always ask
It prepares everything, never sends without a click.
Everything sits there ready; only what you approve gets sent. Good for the first weeks.
-
Balanced
The default. It sends what it has safely mastered.
Confirmations and routine to known contacts it sends itself. Anything new or delicate it puts in front of you — with a draft and a reason.
-
Autopilot
It acts on its own, except in five cases.
It sends by itself — except for uncertainty, new contacts, money, legal matters and cancellations. These five exceptions cannot be switched off.
It proposes changes itself. After five approvals waved through unchanged: “Should I send this directly from now on?” After two corrections in the same place: “With this kind, I’d rather ask from now on.” What runs on its own at which level is in the table.
The breakdown
What runs on its own at which level.
Five actions, three levels, nothing left to interpretation. The last two rows require approval at every level — they cannot be switched off.
-
What that means
- Always ask
- It reads, sorts and prepares — nothing leaves your device without you having seen it.
- Balanced
- Whatever happens only on your device, it does. The moment something goes out, it puts it in front of you.
- Autopilot
- Familiar work runs through without you — with the two exceptions that always hold.
-
Read and classify mail
- Always ask
- on its own
- Balanced
- on its own
- Autopilot
- on its own
-
Send appointment confirmation
- Always ask
- approval
- Balanced
- on its own
- Autopilot
- on its own
-
Mail to a known contact
- Always ask
- approval
- Balanced
- approval
- Autopilot
- on its own
-
First mail to a new contact
- Always ask
- approval
- Balanced
- approval
- Autopilot
- approval
-
Delete for good · move money
- Always ask
- approval
- Balanced
- approval
- Autopilot
- approval
Deleting and payments need your approval at every level — until you decide otherwise, item by item.
Choosing the model
Every model gets one real call — before it is offered at all.
The gateway’s catalogue is not proof: it lists models that reject every request with an error. A server set to one of those answers every question with “I can’t reach the language model right now.” So here things are tried out, not taken on faith.
The line runs from your device via the server to the thinking service.
-
Fetch the catalogue
The gateway names its models. At the last measurement there were eighteen — including names that cannot actually be used there.
-
Call every one of them
Each model gets one short, real call: one after another rather than all at once, and under both name forms a gateway may accept. Nine of the eighteen actually answered.
-
Only what answers and thinks
Only what answered and delivered a line of reasoning is offered. Models that answer without thinking are not on the list: without that reasoning, what reached the user was “The thinking service answered but said nothing”.
- 18 models the catalogue names
- 9 actually answer
- 24 h a result holds, then it is rechecked
State of the last measurement at the gateway. A result holds for one day, then it is tried again — each check is eighteen small calls, which is why it is not done more often.
In normal operation, Claude Sonnet 5 does the thinking for you. If Anthropic cannot be reached, DeepSeek V4 Pro takes over — a provider based in China. Where both sit and what is transmitted is set out on the security page.
When the computer was off
A missed schedule is caught up, not hushed up.
The scheduler lives in the database on your device, not in the operating system. That is how OCTOVO can work out after every start what fell due while it was off.
The day runs from the overview via mail and appointments to the end of work.
-
One beat every minute
A beat checks what is due. A new one waits for the previous to finish — otherwise a morning briefing that takes three minutes would run three times over.
-
The gap is measured
If more than five minutes lie between two beats, OCTOVO looks for the due dates that fell in between.
-
It says so before it acts
“… should have run at 7:00 — I’m catching it up now.” The sentence appears in the conversation before anything happens.
-
The log says “caught up”
A caught-up run is labelled as such — in the order and in the log. OCTOVO does not pretend everything was on time.
With the emergency stop engaged, no beat runs at all. And drafts that would have landed in your quiet hours only go out once your working hours have started again.
The memory
What it remembers and what it forgets on its own.
Every lesson stands as a readable sentence with source and date in a list on your device. There is no hidden personality value and no score describing you.
What is filed expires over time.
What it learns from
- Corrections
- What you change in a draft. One change is enough — afterwards it asks once whether it should remember.
- Decisions
- What moves into a folder, what you delete, what you forward. The rule comes from the doing.
- Observation
- How you write yourself: sentence length, sign-off, favourite words. It reads your own mail, not other people’s.
- Direct instruction
- “Remember: Ms Kern can only be reached in the morning.” The most direct way — and it confirms in one sentence what it understood.
What disappears again
- Faded after 90 days
- A memory that has not been confirmed in 90 days and never came up again no longer counts. It is not deleted — it is put to you for confirmation at the next fitting moment.
- On a single word
- “Forget the last hour” deletes immediately — what was read as well as its search index.
- Never learn again
- A blocked lesson does not come back, even if the same trigger occurs a second time.
- Tidying with the beat
- The scheduler measures the memory as it goes, merges duplicates — and tells you when a limit kicks in.
Memory lives in the same encrypted file as the rest of your data. We have no copy, and it never goes to the language model in one piece.
Proof
Every action with trigger, reason and result
Seven columns per row. The log is the reason it can be established after the fact what happened — and why.
- Time
- to the second
- Trigger
- which order, which event
- Action
- what exactly was done
- Data used
- which excerpt was needed
- Reason
- why it decided that way
- Result
- succeeded, failed, declined
- Cost
- what it took from the quota
The log can be filtered and exported — your evidence under Art. 5 Abs. 2 DSGVO. Nobody can delete it.
It also records what did not work: aborted attempts and the cases where it wasn’t allowed.
The hard edge
One press stops everything.
Five things it neverdoes.
They don’t depend on level or switch. They are built in.
-
Delete for good
What it clears away sits in the trash for 30 days, and it cannot empty that. Only you delete for good.
-
Move money
Transfers, purchases, cancellations that cost money: always your hand. Even topping up the quota has a limit you set yourself.
-
Send from the screen
Whatever arises from screen content it has read requires approval. Even on autopilot, without exception.
-
Bulk mail
More than 20 recipients per message or more than 50 outgoing messages per hour won’t go — regardless of any setting.
-
Act in someone else’s name
No invented senders, no borrowed identities, no bought address lists.
Above it all, the emergency stop: one click on the menu bar icon halts every automation. Nothing is lost, everything waits.