Auxten Wang
@auxten · auxten.com
- 🏢 Technical Director @ ClickHouse
- 🚀 Creator of chDB — acquired by ClickHouse 2024
- 🧬 Maintaining a pile of side projects… alone
- 📝 Shipping 3 iOS apps as a one-man company
- Ex-Shopee · CovenantSQL · Baidu · Qihoo 360
How We Got Here
Four years of AI coding, compressed. It moved absurdly fast.
- 2022 ChatGPT. Describe a function, copy the answer, paste into your editor.
- 2024 Cursor + Claude Code. The AI moves inside the editor — reads your repo, runs your tests.
- 2025 CLI agents & multi-agent pipelines. Codex and Claude Code in the terminal: test → fix → review → benchmark.
- 2026 Autonomous, always-on agents. The question stops being "can it code?" and becomes "can it run without me, 24/7?"
From Copilot to Coworker
Same models underneath. Completely different relationship.
IDE Copilot
- You drive, every keystroke
- Lives in your editor session
- Forgets everything when the tab closes
- Stops the moment you close the laptop
Autonomous Agent
- Lives on a server, runs 24/7
- Persistent memory across sessions
- Reaches you on chat — Discord, Telegram…
- Learns its own skills, gets better over time
Think of it as hiring an ops teammate, not buying a faster autocomplete.
What is Hermes Agent?
Open-source autonomous agent by Nous Research · MIT · shipped Feb 2026 · ~195k★ on GitHub
Not a chatbot. Not a copilot. An agent that lives on your machine and gets smarter every day.
My Use Stack
Deliberately boring. One chat app in, the agent does the rest.
No new SaaS, no per-seat fees. Recurring cost: ~$200/mo — the Codex subscription I already pay for.
Migrating My Brain in One Sentence
A year of teaching Claude Code, Codex & OpenClaw my skills and conventions. Moving it all to Hermes took… one message.
~/.openclaw/skills/ to ~/.hermes/skills/openclaw-imports/.Verified: source 27 → target, no mismatch;
skill_view loads fine. New session: /reload-skills and you're set.It also reads my CLAUDE.md / AGENTS.md — the conventions I taught Cursor and Codex just carry over.
Why I Trust It — #1
It never silently drops a task
- Instant ack. The second I send a task, it drops a ✅ on my message — before any LLM call. Backend model flaky? Doesn't matter; I know it landed.
- Closed loop. Every message gets a final result. Success or failure, it always comes back and tells me.
- No more "did it even get my message?" No more staring at a dead chat.
17 minutes later — it came back, like it always does.
Why I Trust It — #2
Background cron that actually runs my ops
| Schedule | What it does (and reports to Discord) |
|---|---|
| Sun 08:00–09:30 | Weekly Apple Search Ads optimization for 3 appsper-app keyword/bid tuning, then one roll-up report |
| Daily 08:00 | Cross-app anomaly inspectioninstalls / trials / revenue / LLM errors — read-only, pings me only if something's off |
| Daily 10:00 | Emby "pick of the day"one title, tuned to my taste |
| Every 14 days | App Store keyword + metadata refresh, then ship a new version |
| Sat 15:00 | Download the week's new anime episodes to my NAS |
I optimize ad spend every week and inspect every online task every day — without lifting a finger.
Why I Trust It — #3
Hermes as my foreman (监工)
cc-supervisor; the dashboard refreshes every pane every 30s. Pane 0:0 (Claude Code) is idle at the prompt with commit typed but not run — I won't press Enter to commit on my own unless you explicitly ask.An agent… supervising other agents. And it knows where to stop and ask.
Use Case 1: Analyzing My Own Apps
I ship 3 iOS apps solo. Ads, users, revenue, releases, crashes — scattered across 6 dashboards. So I stopped opening dashboards.
Ad spend, user activity, releases — one agent, one place, plain language.
It Doesn't Just Dump Numbers
Yesterday: spend $106.63 · installs 32 · CPI $3.33
Last 7d: spend $440.16 · installs 104 · CPI $4.23
Since re-enabling: 6/14 2 → 6/15 27 → 6/16 32 installs — clear recovery ✅
Weekly ad optimization, per app
One app's GA4
engagedSessions yesterday 10, 7-day avg ≈ 35.8, ↓ ~72%; but sessions / users / events all normal → likely a tracking artifact, not the app being down.Daily anomaly inspection
It tells me whether to panic — and usually, not to.
Use Case 2: What Should I Watch Tonight?
Not everything has to be work. I taught it my taste once…
…and every day at 10am it gives me one pick from my own Emby server — with a reason, and never something I've already seen.
One Pick a Day, With Taste
Peels back the drug trade, the police, politicians and the press, layer by layer — the criminal ecosystem of Breaking Bad + the power games of House of Cards. S01E01 is on your Emby.
An ordinary man swallowed by a criminal system, with Billions-style money games — colder and more oppressive.
Bonus: every Saturday it grabs the latest episodes of my favorite cultivation anime and uploads them to my NAS. A feedback loop, for my couch.
First, the Memory Trap
Build on my laptop (Cursor / Claude Code / Codex) → hand off to Hermes on the mini. Context evaporates.
The obvious fix: "just remember everything." That's the trap.
Auto-capture pollutes: a one-week vegetarian trip becomes my "permanent diet"; a Firebase idea I dropped months ago still haunts every design.
Bad memory is worse than no memory.
ClickMem — a memory safe
Remember what matters, forget what stopped being true, never treat a passing detail as your identity. A safe, not a firehose.
- Explicit-only. Two ways in: an agent commits a refined conclusion (
clickmem_remember), or I import curated docs (AGENTS.md/CLAUDE.md/ rules). Raw chat stays cold — never auto-injected. - Conflicts surfaced, not hidden. A new memory that contradicts an old one gets flagged for me to resolve — not silently stacked.
One Local Server, Zero Server-Side LLM
Embedding-only recall, scoped by project / privacy / source · chDB or ClickHouse.
The Button Software Couldn't Click
Hermes pushed an update → a pile of macOS permission dialogs.
macOS explicitly prevents software from clicking security dialogs.
The agent that controls my computer… can't click "Allow" to control my computer.
💡 What if I use hardware?
$182 Rock 5B + IP-KVM
- USB HID emulation (keyboard + mouse)
- HDMI input — real video capture
- RK3588 hardware video decoding
- Agent helps agent: mutual passwordless SSH — if one dies, the other revives it
Total system draw: 7 watts
Open-sourced as HandsOn — one MCP interface, many backends (macOS, Rock 5B, PiKVM, NanoKVM).
Submitting a PR to Itself
- My agent found a bug — in its own platform's code.
- Plugin reload → restart loop → gateway death.
- It diagnosed root cause, wrote a one-line fix + 2 tests.
- Opened the PR under its own GitHub account.
- Three human reviewers approved it.
An AI agent. Finding a bug in the platform it runs on.
Submitting a patch. Approved by humans.
The Stack Today
- Daily dev: Cursor + Claude Code + Codex
- 24/7 agent: Hermes Agent on Mac mini
- Comms: Discord (phone + desktop)
- Backend LLM: my personal Codex sub + OpenRouter MiniMax M3 backup
- Memory: ClickMem — chDB + Qwen3
- Hardware: HandsOn IP-KVM (Rock 5B)
Recurring cost: ~$200/month — the Codex subscription I already pay for.
Five Things I'd Tell You
- Infrastructure > prompts. — System engineering is the moat.
- One sentence can migrate your whole brain. — Portable skills + memory beat any lock-in.
- Bad memory is worse than no memory. — A real partner has to be able to forget.
- Reliability earns trust. — Instant ack + a closed loop beat raw capability.
- Treat it like a coworker, not a tool. — Containment = safety + freedom.
Keep Building,
Keep Fresh.
Hermes Agent · chDB · ClickMem · HandsOn