Agencies that build their own tools tend to build the same five. Each one replaces a task someone does by hand today. You describe the tool to Claude, ChatGPT or Cursor, and it writes the code against Build's data.
This guide covers what each tool does, what it needs and a good order. Build them in this order, because each one teaches you something the next one uses.
What you need
- A Build account at build.sellermate.ai with client accounts connected in the console.
- An API key from the console. Keys look like
sm_live_.... Use read-only keys for everything except the one write step in tool two. - An AI assistant, and a place to run scripts on a schedule.
Start every prompt the same way: "Read the API docs at https://www.sellermate.ai/api-docs first. Call get_user_context first to list the accounts."
1. The Monday client report
One Slack message per client each Monday. It shows spend, sales, ACoS (ad spend divided by ad sales) and orders against the week before, plus the three campaigns that moved most.
It is the task agencies repeat most, and clients read it. It uses a read-only key and takes about 30 minutes. The full steps are in the guide "Send a Monday client report to Slack".
2. The wasted-spend finder
A list of search terms that spent money in the last 30 days and sold nothing, with a total. It adds them as negative keywords only when you run it with a flag. A negative keyword tells Amazon to stop showing an ad for that search.
The audit needs a read-only key. Applying negatives needs a key that can write, so limit that key to one account. It takes about 20 to 30 minutes. See "Find wasted ad spend".
3. The budget pacing alert
Every hour during working hours, the tool projects each campaign's spend to the end of the day. It posts two short lists to Slack. One is campaigns likely to run out of budget early. The other is campaigns spending far too little.
A daily report cannot catch a strong campaign that stops at 2pm. This can. It uses get_campaign_performance with a read-only key and takes about 30 minutes. It only helps if someone watches the channel, so send it to the people who can act.
4. The stock-out guard
Each morning, the tool works out how many days of stock are left for every advertised product. It warns you when a product will run out while ads are still spending on it. Aim for a warning about two weeks ahead.
Ads on a product you cannot ship waste money, and a long stock-out can hurt the listing's ranking. This tool needs a Seller Central connection for inventory (get_inventory) and campaign spend. It is a read-only tool and takes about 30 minutes.
5. The client dashboard
A private web page for each client with five key numbers, a daily chart and two top-ten tables. It is the biggest of the five at one to two hours.
Build it last. Once the Monday report and the alerts exist, clients ask where they can look for themselves, and you know which numbers they care about. Use a read-only key limited to that client's accounts.
Check it
For each tool, run a dry run first. Compare one account's numbers with the console or Seller Central for the same dates. If they differ, check the date range and time zone before anything else.
Run it every week
The report is weekly. The pacing alert and stock-out guard run on a schedule you choose. A small server, a GitHub Action or a Vercel cron job is enough. The wasted-spend finder works best weekly too, with the CSV posted to a channel for a person to review.
Watch out for
- Keep writes behind a flag or an approval. Build adds its own checks on top: budget caps, preview before apply, idempotency keys and an audit log. A blocked write returns a 422 error and nothing reaches the marketplace.
- Do not rebuild what you can ask for in chat. A one-off question is faster in Claude or ChatGPT. See "Chat or automation".
- Time estimates are rough. Your first tool takes longest.
- Rate limits are 60 requests a minute on Free and Build, 300 on Scale. Many clients in one script may need pauses. A 429 comes with a
Retry-Afterheader. - Savings vary by account. Measure your own wasted spend before you promise a number to a client.