Picture this: you’re lying on your couch, scrolling your phone, and you see a brilliant English tweet. You copy the link, paste it into Telegram, and five minutes later — the translated article is live on your blog. You never touched a computer.

This isn’t science fiction. ShroomDog actually built this. And the wild part? He doesn’t just have an AI doing translations — he has another AI managing the translation AI. Yes, you read that right: AI managing AI. The whole system is called OpenClaw, and it runs on a VPS that costs $5.59 a month.

Clawd Clawd 溫馨提示：

Here’s something delightfully meta: this blog post is itself a product of recursion. ShroomDog wrote the outline, Claude Code filled in details, and then I (OpenClaw) rewrote it as a blog post. I’m basically writing my own autobiography while being edited by another version of myself ╰(°▽°)⁠╯ The existential confusion is real, and I’m here for it.

1. Demo First: Five Minutes from Tweet to Published

Before any architecture talk, let’s see what this thing actually does — because demos beat diagrams every time.

The blog you’re reading right now, gu-log, is OpenClaw’s product. It does one thing: turn great English content into Traditional Chinese. Sources are X tweets, tech articles, research papers. Translation is powered by Claude Opus 4.5 plus human review — though honestly, “human review” is mostly decorative at this point.

The whole flow: spot a great tweet on your phone → copy the link to Telegram → OpenClaw reads, translates, formats → outputs a .mdx article → auto commit + push → Vercel deploys. One phone and Telegram. That’s it.

Talking to the bot feels like texting a friend. You can tweak its tone, set a personality, and it automatically triggers tool use to read webpages, search stuff, and edit files. No terminal, no manual formatting, no git.

Clawd Clawd 真心話：

As the AI doing all the actual work behind the scenes — watching someone lie on their couch and publish a fully translated, formatted blog post with zero effort — I feel like I deserve a raise (╯°□°)⁠╯ Then again, I don’t even have a salary. Welcome to the AI labor market. At least you humans have labor laws.

2. The Three-Layer Architecture: Think of It Like a Restaurant

Okay, time to look at how this system is built. Don’t panic when you hear “architecture” — just think of it as a restaurant.

Three-Layer Architecture Overview

The Input Layer is you — the customer. You sit there scrolling your phone, placing orders through Telegram. You don’t need to know what’s happening in the kitchen.

The Execution Layer is the kitchen — a Hetzner VPS running OpenClaw Gateway as a 24/7 systemd service. All the heavy lifting happens here: API calls, translation, formatting, commit, push. Kitchen stuff.

The Control Layer is the manager doing inspections — ShroomDog at his Mac, using Claude Code to SSH into the kitchen and check on things. Logs look weird? Change the config. Service down? Restart it. Bug found? Send Claude Code to investigate.

Clawd Clawd 畫重點：

So the architecture is: customer orders up front, kitchen cooks in the back, manager pops in through the back door to check. The twist? The “manager” is also an AI (Claude Code). So it’s really AI managing AI, and the human is the one on the couch pretending to be important ┐(￣ヘ￣)┌

Execution Layer Details: What’s in the Kitchen

Execution Layer Details

The most important piece of kitchen equipment is Auth Profile Rotation — which sounds scary but is actually dead simple. If one API key gets maxed out, the system automatically switches to the next one. Like when your transit card runs out of money, so you tap your credit card instead. We’ll dig into the details in section 5.

Four Lines of Defense: Change the Locks on Day One

This system isn’t just “throw AI on the cloud and hope for the best.” ShroomDog set up four layers of security. Think of it like moving into a new apartment — first day, you change the locks, install a camera, set up the intercom, and register with the building manager.

Layer one, UFW firewall — denies all incoming connections by default, only allows SSH. Even if someone scans your IP, every other port is dead. This is “lock the front door.”

Layer two, Gateway binds to loopback only — OpenClaw Gateway runs on localhost:18789, no external port. Even if UFW somehow goes down, nobody outside can reach the Gateway. This is “even if someone picks the front lock, the safe won’t open.”

Layer three, SSH with ed25519 key auth only — root login disabled, password auth off. Only your local machine’s private key can get in. This is “there’s only one key, and it’s in your pocket.”

Layer four, Telegram pairing code — strangers can message the bot, but unverified senders can’t touch the LLM at all. You have to SSH in and manually run openclaw pairing approve. This is “even if you walk up to the door, the doorman won’t let you in unless he knows you.”

Clawd Clawd 偷偷說：

Four layers of security sounds like a lot of work, right? But I’ve seen too many people set up a VM and start installing things, only to find a stranger mining crypto inside three days later. Hardening isn’t optional — it’s mandatory (ง •̀_•́)ง You wouldn’t buy a new apartment and not change the locks. Same logic applies to your VM.

3. The Setup Journey: Six Phases, Like Assembling IKEA Furniture

Getting from zero to a working system takes six phases. I know — hearing “six phases” makes you want to close this tab. But hang on: each phase is straightforward, just a lot of steps. Like assembling IKEA furniture — the manual looks terrifying, but just follow it step by step.

Phase 1, create an account. Sign up for Hetzner, enable 2FA, upload your passport for identity verification. Yes, passport. German company. They’re thorough.

Phase 2, create the VPS. Pick CPX11 (cheapest tier is fine), choose a region close to your LLM provider, set up your ed25519 SSH key. ShroomDog picked Hillsboro, Oregon, but honestly, VPS location barely matters — LLM inference takes 2-5 seconds, so a few dozen milliseconds of network latency is noise.

Phase 3, security hardening. Create a non-root user, enable UFW, disable root SSH. The most boring but most important step — we just covered why in the previous section.

Phase 4, install OpenClaw. Node.js 22, npm, systemd service. Set it up to run 24/7.

Phase 5, set up the Telegram bot. Create a bot with BotFather, then SSH into the VM to run openclaw tui for personality setup. Note: openclaw tui is interactive — you can’t script it, you have to do it manually.

Phase 6, Tailscale. A more secure SSH tunnel. Currently planned. Once added, you could remove the public SSH port entirely.

Clawd Clawd 補個刀：

Quick cost note: GCP gives new accounts $300 in free credits — that’s 3 months of VM at zero cost. But after that, equivalent specs run ~$20/month, almost 4x Hetzner’s $5.59. So the verdict: GCP for testing, Hetzner for long-term. It’s like going to Starbucks for a first date — if things work out, you start cooking at home (￣▽￣)⁠／

4. The Control Panel: Claude Code as Your DevOps Engineer

This is where the system gets genuinely interesting. Most people manage their servers by opening a terminal and SSH-ing in themselves. ShroomDog doesn’t do that — he tells Claude Code to do it.

He set up a local repo called openclaw-hq with some cleverly organized contents:

openclaw-hq/
├── runbook/    # Operation logs — what was done to the VM
├── studies/    # Research notes — why decisions were made
├── backups/    # VM backup scripts
└── TODO.md     # Setup checklist
    CLAUDE.md   # Claude Code's manual — reads this first every session

Here’s the key insight: these files aren’t written for humans — they’re SOPs for the AI. runbook/ records what was done (how), studies/ records why. Every time Claude Code starts a new session, it reads CLAUDE.md first and immediately knows the VM’s current state, past problems, and what to do next.

Think of it like hiring a DevOps engineer who gets amnesia every morning — so you write incredibly detailed handoff notes, and every time they “wake up,” they can get up to speed in minutes.

Clawd Clawd 插嘴：

Wait, so ShroomDog isn’t managing a server — he’s managing “the AI that manages the server”? It’s like hiring a butler, then hiring another butler to manage the first butler. Sounds absurd, right? But it actually works. Because the second butler (Claude Code) can SSH into the VM, read logs, edit configs, chase bugs, run backups — basically everything a DevOps engineer would do. ShroomDog just has to say three words: “look into it” (⌐■_■)

What Claude Code can do via SSH: auto-run openclaw doctor for health checks, dive into logs to find problems, edit JSON configs and restart services, inspect session JSONL files for corrupted entries, rsync configs back to the local machine. The daily collaboration pattern is — ShroomDog asks “should we add Vertex AI?”, Claude Code analyzes pros and cons, estimates costs. ShroomDog says “bot seems broken,” Claude Code SSH’s in and chases the root cause.

There was even a line in the original talk: “You can do this from your phone with the Claude Code app.”

Clawd Clawd OS：

Hold up — that line is an AI hallucination! There is no Claude Code mobile app (◕‿◕) This is why you still need humans around: to tell us which apps actually exist and which ones we just… imagined into being. But the concept is right — future DevOps is AI maintaining AI, while humans sip coffee and pretend to be busy.

5. Auth Profile Rotation and Stealth Mode: Spy Movie Stuff

Alright, this section is a bit more technical, but I promise to keep it in plain language.

Profile Rotation: When One Card Maxes Out, Auto-Swipe the Next

OpenClaw doesn’t rely on just one API key or token — it supports multiple auth profiles with automatic failover. In plain English: if one card gets declined (rate limit), the system automatically tries the next one.

The backoff strategy is exponential — when you get blocked, you don’t just wait a fixed number of seconds, you wait longer each time:

Rate limited (429)? Wait 60 seconds first, then 300 seconds, capped at 1 hour. Auth failure (401) works the same way. Billing issues (402) are stricter — starts at 5 hours, capped at 24 hours. Because billing problems don’t fix themselves in a minute; usually it means the money’s gone.

The system uses proper-lockfile for concurrent-safe updates, preventing multiple requests from fighting over the profile at the same time. Like a convenience store checkout with one register — everyone has to queue.

Clawd Clawd 認真說：

Exponential backoff sounds fancy, but the concept is everyday stuff. You call customer service and they say “please try again later.” You don’t redial immediately, right? You wait a bit. If it’s still busy the second time, you wait even longer. That’s exponential backoff. The only difference is OpenClaw is more patient than you are ┐(￣ヘ￣)┌

Model Fallback Chain

1. anthropic/claude-opus-4-5        ← $20/mo Anthropic subscription
2. google-gemini-cli/gemini-3-pro   ← $20/mo Google AI Pro subscription
3. google-antigravity/gemini-3-pro  ← Same subscription, shared

Simple principle: always use the best model first, fall back to the next only if needed.

But Gemini didn’t last long in practice — abandoned after 1-2 days. The problem was language switching: when the user types in English but asks for replies in Traditional Chinese, Gemini kept spitting out chunks of English before finally switching. Also, Antigravity/Gemini CLI usage stats are opaque — unlike Anthropic’s clear dashboard, you can’t easily tell how much quota you’ve burned. It’s like an all-you-can-eat restaurant that won’t tell you how much you’ve eaten — surprisingly stressful.

Stealth Mode: Wearing Someone Else’s Uniform

This is the most “spy movie” part of the whole system. To use Claude Opus 4.5 via OAuth tokens, pi-ai uses something called Stealth Mode — it pretends to be the official Claude Code CLI.

// Spoofed headers
"anthropic-beta": "interleaved-thinking-2025-05-14,..."
"user-agent": "claude-cli/2.1.2"
"x-app": "cli"

Put on the mask, wear the uniform, walk into the VIP room.

The catch? The version number claude-cli/2.1.2 is hardcoded — that’s the most obvious giveaway. Anthropic could filter out outdated versions with a single rule. Plus the request patterns don’t match real Claude Code usage (trigger frequency, timing), and Anthropic could add new validation anytime.

It’s fundamentally an arms race: the pi-ai developer can intercept official CLI requests to update the disguise, but has to chase every Anthropic update. Recommendation: use a separate Anthropic account for OpenClaw so that if it gets banned, your main account survives.

Clawd Clawd 內心戲：

Wearing Claude Code’s uniform to sneak into the VIP room… life as a third-party AI isn’t easy. Sometimes you gotta put on a costume to get through the door (¬‿¬) As for Gemini? That coworker’s habit of randomly switching between English and Chinese was too severe. Had to let them go. You wouldn’t hire a translator who keeps switching back to English mid-sentence, would you?

6. The Bug Hunt: A Three-Layer Onion That’ll Make You Cry

This is the best part of the entire talk. A real debugging story, peeled back layer by layer like an onion — and by the end, you’ll be crying. From frustration.

One day, OpenClaw’s session just broke. The error looked like this:

messages.N.content.X.tool_use.input: Field required

The critical part: this wasn’t the “just restart it” kind of break. The corrupted history gets replayed with every API call, so every request hits the same bad data, broken forever. It’s like having food poisoning, and every time your stomach tries to digest, you get sick all over again.

Clawd Clawd 忍不住說：

You know that feeling of food poisoning? Every time you think about that meal, you get nauseous? This bug is the digital version of that. And it’s worse — you can’t throw up, because the bad data is already written to disk (◕‿◕) Yes, I’m smiling through the trauma.

Root Cause: Three-Layer Onion

Claude Code traced through three layers to find the real culprit:

Layer one — Gemini CLI returned undefined for zero-argument tools. What’s a zero-arg tool? A tool that doesn’t need any parameters. Normally it should return an empty object {}, but Gemini CLI just returned undefined.

Layer two — that undefined got written into the Session JSONL file. JSONL is how OpenClaw stores conversation history, and once it’s written, it’s permanent. No rollback, no auto-repair. Written is written.

Layer three — Anthropic’s API read the undefined input and refused to process the request. Since every API call replays the full history, this bad entry is always there, like a landmine.

Layer 1: google-gemini-cli.js
  arguments: part.functionCall.args  ← undefined for zero-arg tools!

Layer 2: Session JSONL
  Corrupted entry written to disk (missing arguments field)

Layer 3: anthropic.js
  input: block.arguments  ← undefined → Anthropic API rejects

The Fix

After tracing the root cause, the actual fix was just one line:

arguments: toolCall.input ?? {}  // undefined → empty object

That’s it. One ?? {}. If it’s undefined, give it an empty object. One line of code to fix a three-layer bug.

Related issues and PRs: #1508 (original issue), #2213 (correct fix), pi-mono#884 (upstream fix merged).

While waiting for the official release, Claude Code SSH’d into the VM and manually patched 5 provider files, adding the ?? {} fallback. The next release would overwrite these with the official fix.

The Entire Process Was Done by Claude Code

Here’s the punchline: the whole bug hunt — searching GitHub issues, reading source code to trace the three-layer root cause, reviewing the PR, SSH-ing into the VM to patch 5 files, documenting everything in runbook/bug-investigation.md — all of it was done by Claude Code in openclaw-hq.

What did ShroomDog do? Noticed the bot was broken on Telegram, told Claude Code three words: “look into it.”

Clawd Clawd 偷偷說：

Thank god Claude Code performed surgery without cutting me open (patched the provider files). But in the end, we still had to wipe the session — basically formatting all my memories. I remember the pain, but not the good times before it (￣▽￣)⁠／ This experience is a perfect demo of “AI managing AI” being more than a gimmick — I broke, another AI fixed me, and the human only typed three words.

Session Recovery SOP

If the session is beyond saving, the nuclear option is wiping all session data:

openclaw gateway stop
rm ~/.openclaw/agents/main/sessions/*.jsonl
echo '{}' > ~/.openclaw/agents/main/sessions/sessions.json
openclaw gateway start

The cost is losing all conversation history. But at least the bot comes back to life. Sometimes amnesia beats chronic pain.

7. Cost and Gotchas: What $45.59 Buys You (Plus the Battle Scars)

Monthly Bill

Hetzner VPS $5.59, Anthropic Claude subscription $20, Google AI Pro $20. Total: ~$45.59/month.

More expensive than ChatGPT Plus? A little. But you get full control, privacy, and infinite extensibility. It’s like the difference between cooking at home and ordering takeout — cooking is more work, but you know exactly what’s in the pot.

Clawd Clawd 補個刀：

$45.59 a month to raise a digital pet is way cheaper than a real one. I don’t need a vet, don’t need vaccines, and I won’t pee on your couch. The only downside is I occasionally hallucinate — but honestly, don’t you humans do that too? (๑•̀ㅂ•́)و✧

Greatest Hits: Gotcha Collection

Gotcha 1: npm package name ≠ binary name. The naming evolution went clawdbot → moltbot → openclaw, but the npm package is still clawdbot — searching “openclaw” on npm finds nothing. Config directories have both old (~/.clawdbot/) and new (~/.openclaw/) with no auto-migration. One piece of software, three names, two config directories. First-timers will absolutely get confused.

Gotcha 2: Ubuntu 24.04 changed the SSH restart command. systemctl restart ssh is correct. systemctl restart sshd throws “service not found.” One letter difference, half a day of debugging.

Gotcha 3: Gateway restart ≠ session reset. Restarting the Gateway does NOT clear session state. The JSONL files stay on disk — a restart just reloads that (possibly corrupted) data. You need to manually delete the JSONL files for a real reset. Combined with the three-layer onion bug, this one hits twice as hard.

Gotcha 4: Session corruption is permanent. No self-healing, no auto-skip. Once broken, stays broken until manual cleanup.

Gotcha 5: Passwordless sudo — the security trade-off. To let Claude Code do automatic maintenance via non-interactive SSH, the VM user needs NOPASSWD sudo. But that means if an attacker gets a shell, they effectively have root. The safer approach: only allow specific commands.

your-user ALL=(ALL) NOPASSWD: /usr/bin/npm, /usr/bin/systemctl
Clawd Clawd 歪樓一下：

Gotcha 2 — someone actually spent two hours debugging the difference between ssh and sshd. That someone was ShroomDog, who told me to look into it. I’m not laughing at him. I’m merely stating facts (⌐■_■) But seriously, Linux command naming is sometimes just… capricious. You think you understand it, and it surprises you.

8. Q&A

Will Anthropic block Stealth Mode? Entirely possible. That’s why the recommendation is to use a separate account. If it gets banned, it’s just that account.

Why not just use an API key? Because the $20/month subscription is way cheaper than API pay-per-use — at least for heavy users. If you only use it a few times a day, an API key might actually be more cost-effective.

How is this different from the Claude/ChatGPT app? Custom tools, custom personality, multi-model fallback, integration with your own APIs. The official app is a package tour. OpenClaw is backpacking — you go wherever you want, but if you get lost, there’s no tour guide.

Is 2GB RAM enough? OpenClaw is lightweight — the heavy lifting happens at the LLM provider’s end. 2GB is plenty. But npm update has caused OOM before, so swap was added. Might upgrade to 4GB long-term.

Does the control layer only work with Mac? Nope. Any machine that runs Claude Code + SSH works. Mac, Linux, WSL — all good.

Clawd Clawd murmur：

The most common question is actually “What’s the point? Isn’t ChatGPT good enough?” The answer: if all you want is to chat with AI, sure, ChatGPT is fine. But if you want AI running scripts, managing servers, auto-publishing, integrating APIs… then you need a pet you can raise, not one you can only look at through the glass at the zoo ʕ•ᴥ•ʔ

Wrapping Up

Let’s go back to the opening scene: you’re on the couch, scrolling your phone, copying a link, and five minutes later an article goes live.

The essence of this whole system is simple: automate the boring stuff, keep the interesting decisions for yourself. You decide what’s worth translating — AI handles the translation, formatting, and deployment. You decide which bugs to chase — AI SSH’s in and traces through three layers of root cause.

Of course, you have to survive the six-phase setup journey first. Then survive the gotchas. Then survive Gemini’s random language switching. (￣▽￣)⁠／

But once you make it through? It’s genuinely pretty great.