I spent two days tinkering with it, but couldn't get a small tool to work.
Last January, I used Claude Code for the first time. At the time I was connected to Kimi's model, trying to make a browser extension. Every time it said "it's done," but when I tested it, it didn't work. All sorts of errors, fixing them back and forth for two days, I was about to give up. Continuing to invest wasn't just about failing to solve it—it was also burning through tokens.
Then Claude released Opus 4.6.
I bought a membership on a whim and threw the same requirements at it. Half an hour, from scratch, done in one go. It called the browser extension to check the results itself, simulated user input to test itself, found bugs and fixed them itself, then told me: "You can try it now."
I opened it and looked—it really worked.
That moment gave me a strong feeling: AI capabilities seemed to suddenly cross a line at some point. At least in certain scenarios, it's not just helpful anymore—it's far beyond what you expect from it.
Different Models, Absurdly Huge Gaps
Later this experience happened again.
Using Claude Code plus Opus 4.6 to make a small plugin, helping a friend install a Feishu (Lark) browser automation tool. I struggled for three to four hours; every time it said it was done, but when running there were bugs. I tried all sorts of methods, but just couldn't get it to work.
Later ChatGPT released 5.4. I said, "Fine, your turn." It ran for over two hours, during which I barely intervened, and finally got it through.
Honestly, by then I had no expectations left, thinking this was probably impossible within current AI capabilities. But it just did it.
Same task—one model just couldn't do it no matter what, but switch to another and it worked. The gap between them isn't a matter of degree—it's a matter of can or cannot.
Couldn't Log Into the Computer
Then one day, something happened.
I don't know what configuration I touched while tinkering around, but suddenly my Mac couldn't log in. Enter username and password, press return, wait a moment, snap—back to the login screen. I tried several times, but just couldn't get in.
Before, I might have thought: ask a friend for help? Contact Apple engineers? But thinking about that process—logging into the website, queuing, making an appointment, remote access—and it might not even get solved, just thinking about it gave me a headache.
Then I thought: can I have AI on another machine fix this?
The problem was, I couldn't even open my computer. How would the AI connect?
I started searching through the network, seeing which devices still worked. Looking around, I found one machine previously connected to the Zhipu AI model. The others either didn't have agents or couldn't connect; only this one still worked. Fine, might as well try—it was a long shot.
I used that machine to remotely connect to the problematic Mac and had AI troubleshoot. It checked on one side, while I kept trying to log in on this end, feeding new error messages back to it.
Half an hour later, it found the cause. Some configuration file had issues; it helped me change it back, and I immediately logged in.
The feeling at that moment was specific. When you really need help and the original methods can't help, an AI agent that can connect to your machine—even if it's not the best model—can just get things done.
Becoming More Like a Hacker
After fixing the computer, I started thinking about something else: can I use a coding agent even when I'm not in front of the computer?
I asked Claude Code what to do. It recommended Tailscale—I had no idea what that was before. After setting it up, all my machines connected to a virtual network. Even my phone could connect.
Since then, walking down the street thinking of something, I pull out my phone to connect to my computer and have the coding agent help me work. Actually, remotely controlling AI to do things has been possible for a long time—most people just don't know how.
Later I encountered another problem: when the phone network disconnects or the app exits, the running task breaks. I asked AI again, and it taught me to use tmux to let programs run persistently in the background. After setting it up, tasks can keep running, and I can check progress anytime by connecting in.
That period was quite magical. Remote control, background running, multi-machine networking—I used to think these were programmer things, far from me. I asked AI one question and it was all done. I really felt like I had become the hacker I imagined before.
The Thing I Feared Most
But after getting used to this capability, the thing I feared most changed.
I'm not afraid of the computer breaking. I'm afraid of the agent dying.
Several times, a coding agent on some machine suddenly wouldn't open due to software updates or configuration changes. You've completely gotten used to going to it when there's a problem; when it suddenly disappears, you don't even know how to fix it.
I later figured out a pattern: as long as there's at least one machine in the network with a working coding agent, I can use it to fix other machines—just like fixing that Mac that couldn't log in before.
But what if the last agent dies too?
Honestly, that would really make me panic.
An Idea
Thinking about this, I realized the problem wasn't just mine.
More and more friends around me are starting to use AI tools and install coding agents, and when they work well, it feels great. But opening channels and configuring environments—these prerequisite tasks are much harder than imagined. Once something breaks, most people don't know how to fix it.
You were enjoying the convenience, and suddenly it stops working. This feeling of disappointment is worse than never having used it at all.
I wondered: can we stop making everyone fiddle with configurations themselves? When you need help, open the terminal, type one command, and have an AI connect to help solve the problem?
This was the starting point for building Aima Service.
Aima Service
I spent time with the team turning this idea into a product.
Several AI agents run in the backend, on standby 24/7. You don't need technical knowledge in advance—open the terminal with one command and it starts.
For example, if the coding agent won't open, it helps you fix it. Want to install a new AI tool or programming environment? Hand it over. Device has strange technical failures? Let it troubleshoot.
We don't guarantee 100% success. AI capabilities are still growing; some scenarios that can't be handled today might work later. But the current success rate is higher than most people think.
We're offering large amounts of free credits now—no need to pay, just log in and use it.
AI's Most Moving Moments
Finally, something unrelated to the product.
My wife doesn't usually use AI much, and doesn't really know when to use it. Once she had a tricky situation communicating with her child's teacher and didn't know how to phrase things. She tried asking DeepSeek. It gave some communication advice, helped her avoid several phrasings that could easily cause misunderstandings, and even encouraged her a bit.
She told me: AI is much more powerful than she imagined.
Since then she started actively seeking AI's help.
I've seen similar changes in myself, in her, and in many friends around me. People's views on AI often change not because of daily convenience, but at moments when they really need help and other methods don't work—they're moved.
I want more people to have such moments. That's what Aima Service is doing.
Go to aimaserver.com and give it a try.
Originally published at https://guanjiawei.ai/en/blog/why-i-built-aima-service
Top comments (0)