How My Hand-Rolled Claude Code Client Stays Completely Ban-Proof

Every few days someone comes and asks me whether there’s a foolproof way to guarantee Claude Code never gets your account banned.

Honestly, there is. It’s just that this approach is a little “inhumane” and costs a fair bit extra, so I never talked about it publicly. But so many people in the group have been asking lately, and on top of that I’ve been using my own web-based Claude Code throughout my previous videos, which has left a lot of people curious about what it actually is. So in this episode I’ll walk through the whole approach, as far as I can actually talk about it.

To be clear up front: most people won’t need this episode, and people just getting started may not follow all of it. But if you’re seriously treating CC as production capacity, this thinking is worth a look.

First, what the web version looks like

I already had this web version built back in February, and after several rounds of iteration it now looks roughly like this.

The core area is a chat window, very much like a Claude Code or Codex client — it hides all the complicated command execution, thinking traces, and code display, leaving only the conversation between you and CC. It turns options into a clickable interactive interface, and renders images and reference documents you drag in as cards. Above the input box it shows, in real time, the current model, thinking depth, context usage, and both my 5-hour and weekly quota burn. You can switch quickly between sessions, and if you want a terminal, you can just open one and type commands.

I also built it a file system. Click in and you can see the files in all my project folders, including AI-generated images and web pages, all browsable and downloadable, or you can generate a share link to send to a friend. There’s also a daily todo section, and a mailbox section — all my domestic and overseas email lands here, and it manages all my Claude Code accounts at the same time.

More importantly, this system isn’t just for me. Other people on my team log in and each of them has their own Claude Code account bound, plus their own virtual host. It’s a multi-user system.

Taking it apart: five parts

Part one is the GUI console server you just saw. Chat, file management, mailbox management. It lets me carry a conversation-driven task across any device without breaking stride — I’m chatting on my computer right now; a minute later I pick up my phone, open the page, and I’m back in that exact task with no seam at all.

Part two: every user actually has their own cloud server, and it’s strictly isolated. Once a user logs into the front end, whether they work through the chat box or the terminal, they’re using the CC on their own server. These servers are physically isolated from each other, and each one has all the fiddly stuff — timezone, IP — carefully set up in advance, so it’s turnkey for the user. To provision a new user, you clone a server.

Part three is a tunnel server, dedicated to establishing and maintaining communication between the front end, the back end, and Anthropic. It monitors connection state, raises alerts on anomalies, and can swap out some of the resources inside at any time. What it connects out through is the part I can’t get into — the residential exit. This way, to Anthropic, every account is just one computer being used normally, within the rules.

Part four: I also stood up a server dedicated to account email. As you know, Anthropic builds a return path into the emails it sends, so I self-hosted a multi-domain mail server and collect the mail from all my domains there. DNS resolution, reply routing, all of it configured by me — clean, and easy to run.

With this whole setup, a user comes in from a browser on any device, passes through the console, and over a persistently held tunnel connects to their dedicated server to operate CC, with all files stored in the cloud. The tunnel server, with its dedicated residential exit, holds a consistent, fixed connection to Anthropic’s servers. That gives you a remote-controlled, absolutely clean setup that also feels good to use.

I even wired it into WeChat, so I can talk to my CC directly through WeChat, and it even helps me run community operations. So when things like Crayfish and Hermès came out and people kept asking whether I used them, I said no — because I’ve always had a better way, wired straight into Claude Code. Why would I go use some third party’s pile of garbage code?

Clean is clean, but how do you pay

This setup only solves clean access. Payment isn’t its job. And there’s a small problem here: most clean residential exits don’t support Stripe payments. Which means you can use it cleanly but can’t complete a payment.

That’s fairly easy to solve too. The two most common methods are gift cards and enterprise team plans. Neither one requires the user to go to Stripe themselves — somebody else pays on their behalf.

Is it worth it?

I put a lot of extra research, work, and money into not getting banned. I think it was absolutely worth it.

Because over the past six months, all the development Claude Code has done for me would have cost at least 400,000 to 500,000 yuan if I’d hired a team, and the timeline would have been more than ten times longer than CC took. What the best tools cost is far, far less than the value they create.

One more thing I’m happy about lately: Kimi’s K3 has, in my judgment, caught up with Fable on front-end work and on taste. And I look forward to a day when we no longer have to use other people’s models this carefully, this quietly, and instead create value with models of our own.

That’s it for today. See you next episode.

Source: EP0066_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] I’m back, friends. I hear some more people got their Claude Code accounts banned recently. But honestly, these days Codex, or WorkBuddy plus Kimi’s K3 model, can already handle a lot of the work. Still, people keep asking me whether there’s a truly foolproof way to guarantee Claude Code won’t get your account banned. There actually is. It’s just a little inhumane, and it costs a lot extra on top, so I never shared it. But lately everyone in the group keeps asking, and on top of that I’ve been using my own web version of Claude Code in my earlier videos — so many people have asked about it.

[00:25] Now, most people won’t need this episode, and beginners probably won’t even follow it. So I’ll try to share my whole approach in a way that’s understandable and that I’m allowed to say. I actually built this web version back in February, and it’s been through several rounds of iteration since. This is roughly what it looks like now. First, the core area is a chat window, very much like the Claude Code or Codex clients. It hides all the complicated command execution, the thinking process, and the code display, and only shows the conversation between the user and Claude Code —

[00:51] just that conversation. It also turns the options into an interactive interface so we can click and choose. And the images or reference documents we drag in get rendered as cards. Above the chat input box it also shows the model currently in use, the thinking depth, and context usage, including the five-hour and weekly quota with a refresh state. You can switch quickly between different sessions too. And if you want terminal mode, you can open the terminal and run terminal commands directly — all of that works.

[01:17] I also designed a file system. Click in and you can see every file in all of my project folders, including AI-generated images and web pages — you can browse all of it right here, which is really convenient. You can download things, and you can generate a share link to send to friends. I also built a daily to-do section and a mailbox section. All my domestic and overseas mailboxes can receive mail in here, and it manages all of my Claude Code accounts at the same time. And a setup like this isn’t just for me. Other people on my team —

[01:42] once they log into this site, each of them is individually bound to one Claude Code account and one virtual host. It’s actually a multi-user system. Let me quickly walk through how the system is built. It breaks into five parts. Part one is the GUI console server you just saw — a chat-style system plus file management and mailbox management. It lets me carry on conversation tasks continuously, in one unified way, on any device. Meaning right now I’m on my computer —

[02:07] using it like this, and the next minute I pick up my phone, open this web page, and it picks up that exact task completely seamlessly. Which brings me to the second thing. Every user actually gets their own cloud server, strictly isolated. Once a user logs into the front end, whether through the chat box or the terminal, what they’re using is the CC on their own server. And those servers are physically isolated from each other, and every server has its time zone, its IP —

[02:32] and all sorts of other complicated settings carefully configured, so it’s turnkey for the user. Setting one up really just means cloning a server. One layer down, there’s a server dedicated to establishing and maintaining the communication path between the front end and the back-end servers, including to Anthropic. It monitors connection status, raises alerts on anomalies, and lets me swap out resources inside it at any time. And what it connects out through is that residential exit we can’t really go into detail about. From Anthropic’s point of view, every account —

[02:57] is just a normal computer being used within its rules. I also stood up a server dedicated to managing the email for these accounts. As you know, Anthropic builds a return path into its emails, so I self-hosted a multi-domain mail server. All mail for all my domains comes into here. How the DNS resolves, how feedback gets sent — I set all of that up myself. Very clean, very easy to work with. With that whole setup in place, no matter which device a user —

[03:23] browses from, they go through this console, and there’s a dedicated persistently held channel that connects them to their own server to operate CC. All the files are stored in the cloud too. The channel server, with its dedicated residential exit, holds a consistent, fixed connection to Claude, to Anthropic’s servers. That’s how we get a remote-controlled, absolutely clean setup that also feels really good to use. I’ve actually wired it into my WeChat too — so, just like Crayfish and Hermès,

[03:48] I can talk to my CC directly through WeChat, and it can even run community operations for me. That’s exactly why, when Crayfish first came out and people kept asking whether I was using it, I said no — because I’ve always had a better way, wired straight into Claude Code. Why would I go use some third party’s pile of garbage code? Now, this setup only solves clean access. The payment part isn’t what it solves. And there’s a small problem here: most clean residential exits don’t support Stripe payments. Which means we can use it cleanly, but there’s no way to complete the payment process.

[04:13] That’s actually pretty easy to solve. The most common methods are gift cards and enterprise team plans. With either of those, the user doesn’t have to go over the network to Stripe to pay — someone else pays on their behalf. So that’s why I never get banned: how much extra research and work I did, and how much cost I put into it. And even so, I think it was absolutely worth it. Because over the past six months, all the development Claude Code did for me — if I’d hired a team to build it —

[04:38] it would have cost at least 400,000 to 500,000 yuan, and the timeline would have been more than ten times what Claude Code took, maybe more. What the best tools cost is far, far less than the value they create. One more thing I’m happy about lately: Kimi’s K3 really has, on front-end work and on taste, caught up with Fable in my judgment. And I look forward to a day when we no longer have to use other people’s models this carefully, this quietly, and instead create value with models of our own. That’s it for today’s video. Bye, bye, see you.