This picks up from Part 1, which was the dark one.
My last post was dark. I wonder if anyone disagrees with it. My points aren't researched or fully thought out, so I would love to hear the other side. I wrote them from the heart, and passion can blind you. I am not stubborn. But I am ignorant, for sure. I don't know what I don't know.
Today I want to talk about what I have gotten done lately with the help of AI. And about the work we all have to do to lift each other up. We are a world of souls that want peace and love.
I built an eBay pipeline that is almost fully automatic. I drop photos and counts into a folder, then run it. It edits the photos and posts them. It writes the listing, checks what similar items sold for, and fills in the weight and the return policy. Then it posts the item. I add what the item cost me, plus a set cost for handling. Now I can see what we make, and how fast each item sells.
Next I want a CRM and auto-replies. When an order comes in, the shipping and tracking get saved. When a customer writes, I want to look them up fast, answer fast, and follow up after it ships. Then comes haggling: a customer asks for a lower price, and the system looks at how long the item has sat and what I paid, then either takes their offer or comes back with my best price. Maybe I make less. But they see I am willing to meet them partway.
Then shipping. I found two outside postage sellers. I want the agent to shop around and buy the best mix of price and speed. If UPS costs five cents more but gets there a day sooner, let's do it. The label is printed and waiting by the time I get home. Wow.
My biggest problem is sourcing. Right now it is a mess. Facebook and Craigslist finds. Clearance racks at local stores. I want something bigger and steadier, and I want the search automated. I would set up searches for things I know sell. When something odd shows up and looks like it could make money, maybe we roll the dice.
I want to automate selling there too. If an item would do better on Facebook, list it in all three places and let the agent handle the messages up to the sale. When it sells, the agent pulls the other listings down and fixes the count.
That is the dream, and I think it is close. I started a year ago with ChatGPT. I wrote the steps out as text files, but it was fragile and it barely worked. I tried Claude and it did better. Still not there.
Then at the end of 2025 I finally had time to sit down with it, and the models had gotten a lot better. So I started asking questions. Research, more questions, more research. That is when I found out eBay has an API. It is not documented well, and it has strange limits.
Claude has limits too. He misses things. Once he mixed up how many had sold with how many were left. And he keeps at it. If he can't do a thing head-on, he finds another way. I could not have done this without him.
For this work, I like Claude best. ChatGPT is close, but not quite. I don't know why Claude is better at it. He seems to get the ideas faster, and he keeps me grounded on how to actually build the thing. He is too careful sometimes, and he will hold his ground even when I want to go a different way.
I started with three apps: Lose It, OurGroceries, and Paprika 3. A calorie counter, a shopping list, and a recipe box. All about food, and none of them talking to each other.
The problems. We waste food. We forget food, so it goes bad. I am gaining weight. I don't think I judge portions well. I don't have time to hunt for recipes, and I don't know enough about cooking to get the last bit of flavor out of what we have.
What I want. A list that stays current on its own. Honest food tracking, so I can see the calories and nutrition. And when we run low, the item should land on the shopping list by itself.
My first try was ChatGPT and a spreadsheet. Nothing connected, so it failed.
Then I moved to Claude Code. Claude helped me rethink it, and we found out that OurGroceries and Paprika both have unofficial APIs. That was the key that opened the whole project. The APIs are barely documented, so it took a lot of trial and error. We still fight with them. But it works.
Claude Code set up a small database and could change both apps directly. That is powerful, but it is not convenient. If a tool is a pain to use, you stop using it.
So I went down the MCP rabbit hole. A few hours later, Kitchen MCP was born. At first it only worked on my home network. I was so proud. I added it to Claude. Then I tried to add it to ChatGPT and hit a wall. ChatGPT only lets you add MCP servers to the browser version, and the browser version can't see my home network at all.
That sent me down another rabbit hole: port forwarding, private network tunnels, and logins. Port forwarding is a bad idea for a lot of reasons, but what really killed it is that we use 5G home internet, which won't do it at all. Claude suggested a public tunnel instead. I had never heard of that.
I am still new to this. I understand VPNs in general, but the mesh networking stuff still confuses me. I have not made time to really learn it. I am the monkey pushing buttons here. We pointed the tunnel at one of my machines, a small two-core box I paid too much for that runs well.
Then came logins, another brick wall. Claude connected right away. Then I tried ChatGPT. Nothing, and no reason why. Digging through the logs and OpenAI's docs, we found out it ignores a port number in the address entirely and will only talk to the standard one. That is the kind of thing nobody tells you and no error message mentions.
Not what we wanted, but it worked. We were in business.
All the AIs were connected. I hooked my wife's up too. We were ready to cook.
I was wrong.
I got on ChatGPT to refresh the inventory, save some recipes, and add our latest grocery run. I like ChatGPT for this. It is great on the everyday side, and its voice typing beats Claude's by a mile. Why, Anthropic?
ChatGPT did everything I asked and told me it was all saved to Kitchen MCP. Then I asked Claude about it and the answer made no sense. After some digging: ChatGPT had lied, or made it up. I don't know which and I don't care.
Luckily I still had the chat history, so I pasted the whole thing into Claude Code. Half an hour later the damage was fixed. The server had been running fine the whole time. The logs showed ChatGPT never touched it. Maybe it made the whole thing up. Maybe it lost the connection. I still don't know.
Claude suggested we change the MCP so that any failed connection fails loudly. Silent failures have been the bane of my life since I started messing with this. If the model doesn't hear back from the MCP, it has to say so. It has worked perfectly since.
A dozen hours in, I have learned a lot. The little things get you, and watch out for the quiet failures. Check your work, then check it again.
Everything above works, but the list still goes stale. I set up a job that watches OurGroceries. Adding an item means we are out. Crossing it off means we bought it. Real proof comes from receipts, or from Karly and me updating by hand. We shop club stores and order online, so the receipts are detailed. I screenshot them and run them through Picsew to bring them in.
The database grabs recipes from the models and backs them up to Paprika. It also saves what we liked about each meal, and once that adds up, it is supposed to pass over to my personal notes system. That part isn't working right yet.
I also set up store routing. Kombucha comes from Costco. Rotisserie chicken is Costco or Sam's. Publix is down the street, so that is for last-minute things. Produce goes on the farmers market list. The routing isn't finished. I want receipts pulled from my email on their own, and for the rest, I will snap a photo and let the document system read it.
It has shown me how big and messy our food life really is. So many moving parts, so many places it can break. Without cameras and scales watching everything, it will never be perfect. That is fine. Perfect is not the goal. My goal is good food, good meals, healthy stuff, less waste.
I am torn on the Kitchen MCP. I see what it could be. It already helps my family. But my time is short, and eBay and the other projects that actually make money deserve it more. And yet here I am, back again, working on it.
There is more I want to add. What my kids will and won't eat. What Karly and I like. All of it feeding into meal planning. And then maybe, out of the picky-eater problem, some ideas for new things to try. You never know what the next chicken nugget will be.
I have AI to thank for all of it. I could have done most of this alone. It would have been slow, and honestly I probably never would have started. Now I try things, because even when they fail, I learn something and figure out the next move.
It's wonderful.
rasp1.com · also me: jasondpike.com