Bu sayfa senin dilinizde henüz mevcut değil — İngilizce sürümü gösteriyor.
Work that happens while you're offline
Tell an agent what to do and when. It runs without you — plans, calls tools, does the research, writes the draft — and it's waiting in your chat when you get back.
Not a notification. Not a summary of what it would do. A real draft, in a real conversation, that you can reply to.
The ceiling: draft, never send
An automation prepares. You send. There is no unattended send, no unattended publish, no setting that unlocks one. Ask for "auto" and the system will tell you it isn't available and set your automation to prepare-and-wait instead. This isn't a default we might change quietly — it's coerced in two separate places in the code, and the background agent's own instructions end with the line "never send, publish, or take any irreversible action unattended."
You create it by talking
There is no builder. No node graph. No cron syntax. You say it, in a sentence, to the agent you want to own it.
Every morning at 7, go through my inbox and draft a reply to anything that actually needs one. Don't send them.
Every Monday, take my newest article and repurpose it into a LinkedIn post, an X thread and a newsletter edition.
That last one matters: the agent is instructed to push back on a cadence it thinks is unwise before it creates anything, and to propose a saner one. It's allowed to disagree with you.
The automation then runs as the agent you asked. Maya drafts the email. Ashley drafts the post. Lucy writes the document. Blake pulls the numbers. Each one arrives with only the tools that agent is allowed to touch — Ashley cannot open your inbox on a schedule any more than she can in a live chat, because the boundary is a runtime access control, not a prompt.
Talking is the only way to create one
The /automations page in the app lets you pause, resume and delete. It does not let you create or edit. Changing the schedule or the instructions means telling the agent — "move my morning automation to 7am", "pause the weekly one", "delete that". We'd rather be honest about this than mock up a builder that doesn't exist.
It runs the real agent — not a template
This is the part most "AI automation" tools skip. A firing doesn't execute a hardcoded script with your prompt pasted into a slot. It boots the same agent loop as a live chat, with the same tools, the same Neural Core memory, and the same draft cards.
Which means the output isn't a report about work. It's the work:
| You asked | What's waiting for you |
|---|---|
| Draft replies to my inbox | Editable email drafts, one per message that needed one |
| Repurpose my newest article | A real document you can edit, export, or send to Drive |
| Prepare a post about X | A social draft card with caption, hashtags and image — ready to publish with a tap |
| Pull yesterday's numbers | A spreadsheet or an analysis, produced from real data |
And because every firing opens a conversation in that agent's chat, seeded with your instruction as the first message, you can just… keep talking. Open it, read the step-by-step timeline of what it actually did, say "make it shorter", and carry on. It's a conversation you weren't present for, not a black box that emitted a file.
You'll know it's there: the agent's avatar picks up a red badge. "Prepared a draft while you were away — tap to review."
The schedule model (small on purpose)
Two shapes. That's all there is today.
- Daily at a time —
07:00, in your timezone, read from your settings. Never asked for, never guessed. - Every N minutes — minimum 30. Anything more frequent is rejected outright.
What we don't have
No cron expressions. No "weekdays only". No native weekly or monthly trigger. No end date. No Run now button. If you want something weekly, the honest version is an interval of 10 080 minutes — it works, but it's a workaround, not a feature. We'd rather you know that before you sign up than after.
The timezone handling underneath is not best-effort. Every schedule is computed with the Temporal API against your IANA zone, so daylight saving and half-hour offsets are correct by construction, not by someone remembering to add a special case. What's stored is always UTC; what the agent says to you is always local — it's explicitly forbidden from quoting a raw UTC timestamp back at you.
What stops it from going wrong
Background work you can't see is only worth having if it can't hurt you. Every one of these is enforced in code, not asked for in a prompt.
It can't send. Covered above. It's the whole design.
It can't run away with your credits. Each automation has a daily firing cap (4 by default), and your whole account is capped at 50 firings a day across every automation you own. Both caps fail closed — if the check itself errors, the run doesn't happen. Every occurrence is deduplicated by a unique key, so a retry or a clock skew can't fire the same slot twice.
It pauses itself when it's broken. Three consecutive failures and the automation disables itself and tells you why — "paused after repeated failures" — instead of quietly burning credits forever. It does not un-pause itself on a later success; you re-enable it deliberately.
Running out of credits is a skip, not a failure. If your balance is empty, the firing is marked skipped — out of credits and moves on. It does not count as a failure, so it can never trip the self-pause. Top up, and your automations pick straight back up. A broken automation should pause itself; a temporarily broke account should not lose its automations.
It doesn't fake success. If the agent announced it would produce something and then didn't, the runner forces the producing tool — up to two attempts — and if it still can't, it has to state the blocker in one sentence. Every run records why it ended: completed, needs input, incomplete, out of budget, or error. Narrating a deliverable it never made is a caught failure mode, not an accepted one.
It doesn't loop. A background run gets an anti-repeat guard (the same tool call with the same arguments is refused rather than re-run), a per-run scratchpad of what's already failed, and a stall detector that after three fruitless turns tells it to either change approach or stop and say what's blocking. Plus a hard wall-clock cap with a graceful final turn, so it wraps up instead of being guillotined mid-sentence.
It keeps its own notes. Within a run, the agent has a durable scratchpad it writes plans, research and draft output to as it goes — so if a run is interrupted and picks back up, it continues from those notes rather than starting the thinking over.
It survives a disconnected app. If a run stalls because an integration dropped, reconnect it and the paused tool call is replayed automatically — within 24 hours, after which a stale intent is allowed to die rather than surprise you.
Approvals can't double-fire
When you tap Send on something an agent prepared, the action is guarded by an effect-level idempotency record. A replay, a double-tap or a crash-after-send can't produce two emails — and if the outcome is genuinely unknown, it's reported as in-flight rather than blindly retried. Unknown is treated as dangerous, not as "probably fine".
Daily routines you can just switch on
Separate from agent automations, three toggles in /automations — no prompt needed:
- Daily social briefing (Ashley, 09:00) — what's scheduled, what performed, what to post next.
- Daily inbox digest (Maya, 08:00) — your inbox categorised into Priority, Work, Bills & Payments, Shopping & Deliveries, Security, Personal. Localised into all 11 agent languages.
- Order status watch (Sam, opt-in, Shopify) — an unread customer email asking "where's my order?" gets the order looked up and a reply drafted, ready for you to read and send. Capped at 10 a day. Drafts only.
That last one shows a rule we hold everywhere: a detector resolves who's actually opted in and connected before it spends anything. If you haven't connected Shopify, nothing runs, nothing is analysed, nothing is billed. We don't question the void.
What it costs
Creating, changing, pausing or deleting an automation is free. No charge for the tool call, no row in your ledger.
A firing costs what it actually used — metered from the real model and tool cost of that run, not a flat "per automation" fee, because a run that reads three emails and a run that writes a 2 000-word document aren't the same thing and shouldn't cost the same. Credits are the unit.
Reads, searches, drafts and plans are free. You're charged for work that produced something.
See Pricing for plans and top-ups.
What an automation knows about you
The stored automation row holds three things: your instruction in plain language, the schedule, and whether it's on. No token, no secret, no message body — ever.
A background run never mints a session for itself. When it needs your Gmail, the token is resolved through a locked-down database function that takes your user id and returns only what that specific call needs — never logged, never written into the job. The automation row itself is protected by row-level security like everything else, cascades on account deletion, and is in the erasure purge list.
More on the model: Security · The Neural Core
FAQ
Can an automation send an email or publish a post without me?
No. Not with a setting, not with a plan upgrade, not by asking nicely. Automations prepare a draft and stop — you tap Send or Publish. The "auto" autonomy level is accepted by the interface then coerced back to prepare-and-wait in two independent places, and the background agent's instructions explicitly forbid any irreversible action unattended. If we ever build true unattended sending, it will arrive as a disclosed, separately opted-in feature — not as a quiet default change.
Can I create an automation from the app instead of the chat?
Not today. The /automations page shows everything, and lets you pause, resume and delete — but creating one or changing its schedule means telling an agent: "move my morning automation to 7am". The empty state in the app says exactly that.
Can I run something every Monday, or on weekdays only?
Not natively. The schedule model is daily at a time or every N minutes (30 minimum). A weekly cadence is possible as a 7-day interval, but there's no weekday selection, no cron expression, no monthly trigger, and no end date. It's a real limitation, not a missing checkbox.
Does Unyo come with ready-made automations?
No. A new account starts with zero automations, and there's no template library to pick from. Every automation exists because someone asked an agent for it in their own words. That's a deliberate trade — you get exactly the automation you described, not the closest template — but it does mean the first one is on you to ask for.
Which agent runs my automation?
The one you asked. Nine of the ten agents can own an automation, and each brings only its own toolbox: Maya has your inbox, Ashley has your social accounts, Lucy writes documents, Blake runs the analysis. (Steve is the exception — he builds sites in a live conversation with you, and doesn't take scheduled work.) The boundary that applies in a live chat applies identically at 3am.
What happens if I run out of credits overnight?
The firing is skipped and labelled "Skipped — out of credits" in your activity feed. It is not counted as a failure, so it can't trip the three-strike self-pause. Top up and the next scheduled run proceeds normally — your automations are still there, still on.
What if the automation keeps failing?
After three consecutive failures it disables itself and shows "paused after repeated failures", rather than retrying forever on your credits. It won't silently re-enable on a later success — you turn it back on when you've fixed the cause.
Can I see what it actually did?
Yes, in full. Each firing is a normal conversation in that agent's chat with the complete step timeline — every tool it called, in order. The activity feed on /automations shows recent runs with their status, and clicking one takes you straight into the conversation that produced it. Then you can just keep talking in it.
Will it wake me up with busywork?
It's told not to. The background instructions include an explicit carve-out: if after checking there's genuinely nothing worth doing — the inbox only holds newsletters and automated alerts — it must say so in one short sentence rather than invent a deliverable. An empty morning should read as "nothing needed you", not as three fabricated drafts.
Start with one
Open any agent and describe the thing you do every morning that you'd rather not.
Every weekday at 8, check my inbox and prepare drafts for anything that needs a reply. Never send them — I'll review.
It'll tell you when it's set, and what time that is where you are.