Mark OomsAt our house, we had the same problem every day. What are we eating tonight? Who has time to go to the supermarket? And every week we'd tell ourselves: this time we're actually going to plan our meals ahead. Never happened.
Then I saw someone on X who had their AI agent do their groceries at Albert Heijn. I thought: I want that too. So I put my OpenClaw agent to work.
What it does
My agent logs into my Albert Heijn account, pulls my purchase history, checks which products are on sale this week, searches recipes on Allerhande that match my taste and cooking time, and fills my shopping list. I get meat and cheese from the local farm, so those show up as separate notes on the list.
Every week I get a proposal with three meals, matched against what's on bonus. I approve, adjust, or reject. Then the agent fills the list in one API call.
What Albert Heijn knows about you
This was the most surprising part. Through the AH mobile API, you can pull your complete customer profile. Diet type, price segment, share of wallet, favorite shopping day. Mine says 90% share of wallet and "Very High" price segment. They also flag categories where they think you might be shopping elsewhere. For me: bread, cheese, and meat. They're right.
The agent uses this profile as a starting point to understand your preferences before you've even configured anything.
How it works
Under the hood there's a CLI tool that wraps the appie-go library. The AH mobile API is the same one the Appie app uses. Some of the GraphQL queries are undocumented, like previouslyBought on product search, which returns everything you've ever purchased.
The agent reads a SKILL.md file that tells it exactly what to do: how to log in, how to propose meals, how to check for deals, how to handle the shopping list. You point your agent at the repo and it figures out the rest.
The login is hacky
Fair warning: the authentication flow requires you to open a login URL in your browser, log in with your AH account, and then fish out an auth code from the browser's Network tab (look for ingelogd.json in the response). It's a mobile app OAuth flow, so the redirect goes to appie:// which your browser can't handle. It works, but it's not pretty.
Vibe-coded, from scratch
I'm a product manager with zero dev background. Everything was built by my AI agent through conversation. I described what I wanted, checked the results, and iterated. The skill that product managers have, clearly defining what you want, checking assumptions, working toward an MVP, turns out to be exactly what you need when building with AI agents.
I think this is going to change our field significantly. The line between coming up with an idea and building it is disappearing.
Try it yourself
The whole thing is open source. If you have OpenClaw running, point your agent at the repo and you should be good to go. If you know how to open the Network tab in devtools, you're already halfway there.