翻訳待ち：The first AI use case that genuinely saves me time (outside of programming)

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

The first AI use case that genuinely saves me time (outside of programming) 29 Aug, 2026 It took me a while, but I finally found one. A tedious task that repeats every week... Grocery shopping 😩 Each time, we have to decide what to eat, turn those recipes into a shopping list, find the right products and fill an online cart. I built a system to automate that process. Actual prompt to plan a meal and order ingredients directly from ChatGPT It started when I asked ChatGPT for recipes that fit my nutrition goals. X/Y/Z% carbs/protein/fat; I don't like this or that... I got good results, but the workflow was not very practical. You have to manually aggregate all the ingredients to create your shopping list, then dig through countless unrelated chat sessions to find the right recipe again while cooking. Not to mention that your phone screen always seems to turn off as soon as your hands are dirty 😬 First, fix meal planning Recipe management is not a new thing. After some research, I chose Mealie to manage my recipes, plan my meals for the week and prepare my shopping list. My current Mealie homepage with recipes. Don't judge :) You can define your weekly menu in the meal planner And it prepares your shopping list The nice thing in 2026: you can connect your AI agent to Mealie through a MCP server. The MCP exposes a collection of functions to interact with Mealie. A subset of the tools available in Mealie MCP I use MCPJungle to manage my MCP servers. It allows you to create tool groups with your own tool composition. Here I selected 41 tools (still a lot!) out of the 68 available in the MCP to reduce its context window footprint. With that, ChatGPT can create recipes in Mealie, plan meals... I also connected the MCP to my agent (I use Hermes) and asked it to remind me about the next meal so that I can be ready to cook it. My Hermes agent nudging me into cooking on Whatsapp This was already great, but there was still one piece missing. Fill the shopping cart We usually order our groceries on our supermarket's website and pick them up a few hours later. An agent can surely do that too, right? Letting an agent control a browser is possible, but can be very slow and consume a lot of context. Also, I'm not comfortable letting it handle my credentials, see my personal info, or place an order by itself. That's where I think a custom MCP makes sense. As Simon Willison explained, MCP can be a safer way to build with agents. The agent never needs access to my supermarket credentials or session cookies. They remain inside the MCP server and none of the tools expose them. You also control precisely what each tool is capable of and what it returns. Unfortunately, my supermarket doesn't provide an MCP server. So I had to reverse engineer the APIs for logging in, searching for items and adding them to the shopping cart. It was not straightforward, as their mobile app is well protected to prevent impersonation and the web app requires a Cloudflare turnstile to log in. This is something I really enjoy doing though (I even developed a tool to help 🤓). I won't dive into this in this post, but please subscribe if you want to know more about it. In the end, I exposed only 4 functions through my MCP server. Letting the agent search the supermarket's entire catalog would give it far too many opportunities to make weird choices. The supermarket already has a "favorites" feature, so I built the MCP around that instead. The agent can only choose among products I already buy. This made matching faster and much more predictable. I self-host the whole thing on a Dokploy instance and connect it to ChatGPT through a secure MCP tunnel. Nothing is exposed to the public web and I keep everything local. I'll detail the deployment too if you're interested. Integration of the two MCP servers through ChatGPT Plugins And it works surprisingly well. ChatGPT creating the lasagna recipe in Mealie and filling the shopping cart And the actual shopping cart 🎉 The part that surprised me most is what I didn't have to build. There's no mapping table between recipe ingredients and supermarket SKUs. I don't tell ChatGPT that "500 g ground beef" corresponds to product 123456. The model gets the ingredient, sees the products available through the MCP and figures out the appropriate product and quantity itself. That's exactly the kind of fuzzy task LLMs are good at. It's also very flexible. You can adapt the ChatGPT prompt to your needs. For example, I explain in mine that I buy the vegetables in a different shop (which doesn't have a website unfortunately) so it knows not to try to match them. Overall, this is a pattern I'll reuse for other automations: expose narrow tools for actions, let the model handle fuzzy decisions, keep a human in the loop for sensitive operations (I'm the one actually clicking the button to submit the order 🤡). Custom MCP servers are neat. Feel free to reach out if you think of a tedious task that would be worth automating. Please subscribe if you want to know more about the reverse-engineering and self-hosting part ⬇️ #agent #automation #mcp