The full story

From empty office to
working team — five acts

Most software asks you to learn it. Kelvyr asks you to manage it — like a small, disciplined, slightly obsessive team. What happens when you hire, what runs at 6am, what they may touch, and every safeguard between them and your money.

Act 1 · HireAct 2 · ConnectAct 3 · They workAct 4 · You approveAct 5 · The office

Hire

60 seconds

Connect

2 clicks, your Google

They work

routines, daily

You approve

Allow cards in chat

You audit

ledger + revert

Act 1

Hiring takes a minute. Seriously — we timed it.

Open the catalog and you'll find twelve specialists — an ad manager, an SEO specialist, a designer, a support agent and more. Each card is a working professional, not a template of promises: real capabilities, a real daily routine, a real playbook. Click hire, give them a name (or keep theirs), and they walk to their desk.

The playbook is the point.When you hire a human specialist, you're mostly paying for their habits: check the search terms every morning, never judge a campaign on one day of data, mind the conversion lag, talk to the boss in money rather than acronyms. We wrote those habits down — with practicing specialists — and hard-coded them. Your employee arrives on day one with a senior professional's discipline, and unlike a human, never has a Friday where they skip the checklist.

Got a job that isn't in the catalog? The Build-your-own simulator interviews you with three questions — What does success look like? How much authority should they have? What rhythm — daily, weekly, or on-event?— and shapes a custom employee around your answers. A property manager built a "listing watcher" this way in under two minutes.

And when it isn't working out? Pause them, or fire them. No severance, no awkward meeting, no knowledge walking out the door — the memory and ledger stay in your workspace.

Marcus Ad Manager

watches every dollar in Google Ads

Hire

Sofia SEO Specialist

hunts page-1 wins in Search Console

Hire

Ivy Designer

daily on-brand posts, real image gen

Hire

Nora Support Agent

clears the inbox, escalates with context

Hire

✨ Build your own

"Watch my competitors' prices weekly and brief me on moves" → a custom employee, 3 questions later.

Act 2

Connecting accounts — and the golden rule

An employee without accounts is a consultant; an employee with accounts is staff. Connecting takes two clicks: you sign in with your own Google account, approve the permissions, and every ads account and Search Console property that login can see appears in your office, ready to be assigned to an employee.

Here's the security model in one sentence: your employees can only ever touch what you can touch.There is no field anywhere in Kelvyr where typing someone else's account ID grants access — connections derive strictly from your own Google login's real permissions. We call it the golden rule and it's enforced in code, tested with forged-request attacks, not written in a policy PDF.

Two more hard rules ride along. One account, one office:a business account that's already connected to another workspace can't be silently claimed by a second one. And disconnection is death: unlink an account and your employees lose access that instant — no cached ghosts.

Assignments are explicit, too. You choose which employee manages which account. Running an agency with several ad managers? Each one gets their own book of accounts, and when a request is ambiguous — "raise the budget" which account? — they ask instead of guessing, because acting on the wrong business is treated as a disaster.

The golden rule

"They can only touch what youcan touch."

Your Google connects → your accounts appear
You assign accounts to employees, explicitly
Typing a stranger's account ID → rejected
An account already in another workspace → rejected
Access after you disconnect → dead instantly

We attack this ourselves: forged account IDs, cross-workspace grabs, replayed requests. The self-test suite has to pass before any deploy ships.

Act 3

The 6am shift you never have to think about

This is the part no chatbot does, and the part that makes Kelvyr feel like staff. Your employees run on schedules, not prompts. Whether you open the office or not, the routines fire: data gets read, checks get run, reports get written, decisions get prepared. Silence from you never means work stops — it just means nothing risky happens.

Every morning, before you wake

Marcus's 10 sentry checks

Each sentry is deterministic code with evidence requirements — not an AI vibe. Findings carry the numbers that triggered them.

01

Budget pacing

Is the month's spend on track — or sprinting toward an overshoot?

02

Junk search terms

Which searches ate money yesterday without a prayer of converting?

03

Conversion collapse

Did customers suddenly vanish while spend stayed flat? (Usually the tracking tag, not the ads.)

04

Budget-limited winners

Which campaign converts great but keeps hitting its ceiling?

05

Quality score drift

Are keywords getting quietly more expensive per click?

06

Ad disapprovals

Did a policy bot silently kill one of your ads overnight?

07

Tracking heartbeat

Is the conversion tag alive and firing at a plausible rate?

08

Device & hour skew

Is 2am mobile traffic burning cash your 2pm desktop buyers earn?

09

Impression-share bleed

How much of the auction are you losing — to budget vs to rank?

10

Week-over-week drift

Beyond the noise: is the real trend up, flat, or down?

M

Mehmet · Ad Manager

Morning check-in — 08:00, unprompted

Morning boss — yesterday brought 46 customers at $38 each (last week: $45, so 15% cheaper). I checked all 10 of my daily sentries: two things need your eyes — 23 junk search terms ate $120, and the Northside campaign is losing 31% of its impressions to budget. Both are on your desk with my suggested fix. Nothing else is worth your time today.
2 decisions waiting10 checks ran

Anatomy of a morning report

Notice what it doesn't say: no CTR, no CPC, no impression-weighted anything. The constitution requires employees to speak in customers and money, because that's what you actually manage. The jargon still exists underneath — ask and you'll get every decimal — but the default is a report your accountant could act on.

Notice the comparison, too. Never "yesterday was bad" in isolation — always against last week, with a judgment about whether the difference is meaningful or just noise. And when a standing request exists ("always tell me about the Riverside account first"), the report opens with it, because memory is permanent.

Findings that deserve action don't stop at prose: they arrive as prepared decisions — an Allow card with the evidence and the suggested fix, sitting on your desk. Which brings us to Act 4.

🧠 Permanent memory

said once, kept forever
Reports at 07:00, lead with RiversideFocus: the certification courseNever touch the brand campaignBoss prefers plain numbers, no %

Instructions land here automatically as you chat — and shape every future report and design.

🧾 Cause & effect ledger

every action, forever
Jul 02

Blocked 23 junk search terms

→ $120/mo recovered

revert
Jul 03

Raised Northside budget +$50/day

→ +38 customers @ $39

revert
Jul 09

Paused keyword "cheap course"

→ CPA −9% this week

revert
Jul 11

Proposed display expansion

✗ you denied — remembered

revert

Mode 1 · Autopilot

They run their own day

Routines fire automatically: dawn reads, sentry checks, morning reports, the designer's daily post, the SEO weekly deep-read. Low-risk grunt work happens on its own; anything money-touching still becomes an Allow card. This is the "I have staff" experience — the office works before you do.

Mode 2 · Delivery

They prepare, you drive

Everything is prepared and handed to you; nothing runs unattended. Perfect for the first weeks while trust builds, or for owners who want to stay hands-on. Flip between modes anytime with one toggle — per employee.

Act 4

The Allow card — power, on a leash you hold

Decisions in Kelvyr happen where conversations happen: in the chat. When an employee wants to change something that matters, a card appears right in the thread — what they want to do, why, and the evidence. Two buttons. Allow executes it on the live account. Denykills it — and the employee remembers the no, so you're not re-litigating the same idea next week.

The button is not a suggestion box. It is the only trigger. There's no code path where an employee spends money, launches an ad, or messages a customer without a card being pressed. We built the entire write-system behind this gate on purpose: the negotiation is conversational, the execution is mechanical.

Then comes our favorite part. After executing, the employee doesn't say "done" because the API returned 200. It makes a second, independent read of your account and confirms the change is actually live — the verification readback. If Google rejected something, you get the honest failure with the reason. Receipts, not reassurance.

And if you regret an approved change? Open the ledger, hit revert. Negatives removed, budget restored, campaign re-enabled — with the revert itself logged, so history never lies.

M

Marcus

● online — working for you

Ad Manager · approval flow

The loop, end to end

They propose

evidence attached

You Allow

one tap, anywhere

They execute

on the live account

They verify

read back from the API

It's logged

ledger + revert

What needs your Allow — and what doesn't

Autonomy for the grunt work. A gate on everything that matters.

Read your data — any report, any range✓ Runs freely
Run daily sentry checks & morning reports✓ Runs freely
Inspect index status / audit a page✓ Runs freely
Generate a design draft for your review✓ Runs freely
Add or remove negative keywords✋ Allow required
Change a campaign budget✋ Allow required
Pause or enable campaigns/keywords✋ Allow required
Launch new ads or campaigns✋ Allow required
Send anything to a customer✋ Allow required
Act 5

Where it all lives: the office

Dashboards make you feel like an analyst. We wanted you to feel like a boss walking the floor. Your team sits at real desks in a living scene — and every pixel of it is state, not decoration.

Bubbles = decisions

A raised hand over a desk is a real pending approval. Click through and decide it in chat.

Screens = live accounts

A glowing chart on a monitor means a connected account being watched right now.

Coffee = paused

Pause an employee and they step away — steaming cup on the desk, screen dark, no routines firing.

Everything else bosses ask

Do I need to know anything about AI or prompts?

No. There is no prompt engineering anywhere in Kelvyr. You talk to employees like you'd message a colleague on WhatsApp — "how did we do yesterday?", "block whatever's wasting money", "make me a post for Friday". The professional discipline lives in their playbooks, not in your phrasing.

What happens the very first day?

The moment you connect an account, your ad manager runs the complete first-day review — every sentry, the full account read — and posts findings within minutes. You don't wait for tomorrow's cron. The first morning report typically contains the most findings you'll ever see, because years of accumulated waste get flagged at once.

Can two employees work together?

They share your office and Aria (your assistant) reads across the whole team, so the daily digest connects the dots — "Marcus recovered $120, Sofia found 3 unindexed pages, Ivy's Friday post awaits approval." Deeper employee-to-employee collaboration is on our roadmap.

What if I ignore the office for a week?

Nothing breaks — that's the point. Routines keep running, reports stack up neatly, and anything requiring approval simply waits. Employees never take silence as a yes. Come back Monday and the office tells you exactly what happened while you were gone.

What can employees NEVER do?

Spend or change anything money-related without an Allow. Touch accounts your own Google login can't access. Claim work they didn't do (verification readbacks are mandatory). Invent numbers. Promise background work that isn't scheduled. These aren't guidelines — they're enforced in code and in the employee constitution.

Is my data used to train anything?

Your business data belongs to your workspace. It feeds your employees' memory and reports — that's it. Disconnect an account and access dies immediately.

What does 'verified' actually mean?

After executing a change, the employee makes a second, separate API call to read the current state of your account, and only reports success if the change is actually there. If Google rejected something, you get the honest failure — never a hopeful "should be done!".

Can I undo something I approved?

Yes. Every applied action sits in the ledger with a revert control. Negative keywords can be removed, budgets restored, campaigns re-enabled — and the revert itself is logged too, so the history stays honest.

Autonomy for the grunt work.
Control for you.

That's the whole design, in five acts. The sixth act is your first morning report.

Walk into your office →