Your store's safety layer
The same method as the book, as a server your agent connects to. The platform gives your agent access to your store, but no way to undo what the agent did. This app is that missing layer: the agent operates the store, but every write is wrapped in a backup, an approval, a verify and a rollback. On sale now for WooCommerce; Shopify opens soon.
14-day trial starts when the store is connected · no card · Free read tier stays
An operating system, hosted
The app is the same two-layer system as the book, but you don't build it. Underneath is the system — the loop, custody, the store's memory, the autonomy dial, the scheduler. On top run the modules, the real jobs of a store. The server enforces the bottom layer for you.
The system
The layer underneath everything, the same for every store.
- ·the loop: explore → plan → confirm → execute → report
- ·custody: backup, confirm, verify, rollback
- ·the store's memory, the autonomy dial, the scheduler
The modules
The real jobs of a store, run on top of the system.
- +catalog, merchandising, inventory
- +orders, support, returns
- +marketing, analytics, being found by AI
Access without custody
The platforms hand your agents more keys every year. Nobody hands you the safe. I built this app after I learned that gap the hard way.
In 2025 I set an agent a small job on a client's live store: add one feature to the theme. I wrote the prompt in a hurry and left the agent running on its own. I saw it start writing to the theme and head the wrong way. Had I let it finish, it would have rewritten the core files of a live store's theme, and I had nothing to put them back: no backup, no snapshot, no rollback path. I hit escape in time. Nothing broke. What saved that store was a reflex and some luck, not a method.
That morning is why this exists. The loop, the backup rules, the confirm gate and the rollback path are the safety layer I built afterward, so it could never again come down to luck and reflex.
Access — what the platform gives
A button labeled “connect.”
- ·a token
- ·a plugin
- ·a green checkmark
Custody — what makes a write recoverable
Four properties, enforced.
- ✓a snapshot before the write
- ✓a moment to read the plan
- ✓a check after the write
- ✓a known path back
The book opens on an invented but familiar case: an operator wired in the platform's official plugin, got a green checkmark, and asked the agent to rewrite one product description. The agent proposed, she read it, she approved. The write went out. Then she looked for the undo. There wasn't one. The admin showed the new description, not the old one. The plugin gave the agent the ability to change the store — not the ability to change it back.
“those two things felt like the same thing when she read the announcement. they are not.”
The platform keeps no old value. When the agent overwrites, the previous version is gone.
Without the explore step the agent invents values instead of reading real ones, and writes outside scope.
A bulk write isn't one transaction. Half the products land in the new state, half the old, and nothing tells you.
A loop the server enforces
Your agent (Claude, ChatGPT or Cursor) operates your store through the server. The server isn't just a connection: it runs every change through a five-step loop and locks the plan once you approve it.
Four rules, enforced
Custody is the four properties that make an agent's write recoverable. Almost no platform ships any of them. The server enforces all four.
The server snapshots the exact fields about to change before anything writes. Without it, the write doesn't go.
The plan is readable in full before the write, and approval is one decision on the whole plan, not a rhythm of clicks.
An independent read of the store after the write, compared to the plan. Not the agent's own log, the store's real state.
Every change leaves a receipt and a way back. An error found months later is a diff, not a reconstruction from memory.
- Locked plan
- Once you approve a plan, it locks. A run that drifts off it stops.
- Audit trail
- Every applied change is logged, so the store keeps its own memory of what happened.
- Always-on runner
- Pre-approved jobs fire server-side on the clock, so your machine doesn't have to be on.
the agent runs, you spot-check
the agent proposes, you read every batch
the agent sketches, you drive
Your store's daily work
Fill missing fields, normalize values, enrich products from a source, in batches with a backup first.
Stage a discount, hero and emails as one plan you read before it fires.
Write descriptions and translations in the store's voice, with no invented product data.
Reads and reports can run on a schedule unattended. On writes you keep your hand in.
Restock proposals and the exception queue sorted, a proposed action per row, you approve.
Structured product data and llms.txt in order, so the buyer's AI finds the store.
Pick your path
Priced per store, per month. You connect with no card, and a 14-day trial of the full product begins. The card comes only if you keep a paid plan. Free stays free as the read tier.
One operator, writes on, the whole loop. The most popular start for a single-store operator.
Team: 5 seats, shared approvals, a per-operator log. Business: unlimited, the whole portfolio, with the volume discount.
| Stores | Solo | Team | Business |
|---|---|---|---|
| 1–3 (full) | 49 € | 99 € | 199 € |
| 4–10 (−20%) | 39 € | 79 € | 159 € |
| 11+ (−35%) | 32 € | 64 € | 129 € |
- 1 · Connect
- You connect your WooCommerce store. A 14-day trial of the full product starts at once, with no card.
- 2 · Try it
- You run the whole loop for 14 days: explore, plan, confirm, execute, report, every guardrail on.
- 3 · Keep it or stay Free
- After the trial you pick a paid plan to keep write access (card goes to Stripe then), or downgrade to Free (read-only) anytime in-app.
WooCommerce now, Shopify soon
Pick a tier, connect the store and start using it right away. The 14-day trial starts at connect, with no card.
The Shopify version needs the App Store review, which is in progress. The list is Shopify only — WooCommerce is open and selling now.
The guardrails lower the risk
They don't remove it. An agent under time pressure can cut a corner. That's why you keep your hand in on writes: you read the plan, you watch the first batch land, you stop the agent the moment it steps off the process. Reads are low-stakes and can run unattended. The store is yours, and so is the decision to let an agent write to it.
Common questions
WooCommerce store: yes, right here, today. Pick a tier, connect the store, start using it. Shopify store: opens when the App Store review clears, so join the list and I'll tell you the moment it does.
No. You connect your WooCommerce store and a 14-day trial of the full product starts with no card. The card goes to Stripe only if you keep a paid plan after the trial.
You pick a paid plan to keep write access, or downgrade to Free (read-only) anytime in-app. Nothing charges on its own before you choose a plan.
Each store is priced by portfolio size: 1–3 full price, 4–10 −20 %, 11+ −35 %. The per-store price drops as the portfolio grows.
No. Reads can run on a schedule unattended, but on writes you keep your hand in: you read the plan, you watch the first batch. The guardrails lower the risk, they don't remove it.
Every write has a backup before it and a verify after it. A bad batch is undone from a known rollback path, not from memory.
The hosted version of the same method
The book and the free course teach the loop and the four rules. The app enforces them server-side, for when you'd rather not build them yourself.