How I Manage My
One-Man Company
with OpenClaw 🦞
Auxten Wang · Technical Director @ ClickHouse
Singapore OpenClaw Meetup · March 2026
What is OpenClaw? 🦞
Imagine having a personal assistant that lives on your computer, works 24/7, and never asks for a raise.
- You text it on WhatsApp, Telegram, Slack — it texts back
- It can browse the web, write code, send emails, manage files
- It remembers what you told it last week
- It runs on your machine — your data stays with you
Open source · 309k GitHub stars · macOS / iOS / Android / Linux
Today I'll share what happens when you actually let it run your company.
Thanks to Our Sponsor
The leading open-source OLAP database for real-time analytics.
Petabyte-scale data, millisecond queries. 46k+ GitHub stars.
Powering AI infrastructure for
"ClickHouse played an instrumental role in helping us develop and ship Claude 4." — Anthropic
- 🏢 Technical Director @ ClickHouse
- 🚀 Creator of chDB — acquired by ClickHouse 2024
- 🛒 Ex-Principal Engineer @ Shopee
- 🔗 Cofounder & CTO, CovenantSQL
- Baidu · Qihoo 360 · eLong · 4Paradigm
The Problem
One person. Too many projects.
Code, social media, App Store, infrastructure, community...
I needed a 24/7 coworker.
AI Coding Evolution
- 2022: ChatGPT → copy-paste into editor
- 2023: Describe entire files, better prompting
- 2024: Cursor + Claude Code — AI inside the editor
- 2025: Multi-agent pipelines (test → fix → review → benchmark)
- 2026: Can they run without me? 24/7?
chDB DataStore: Multi-Agent Pipeline
Test generator → bug fixer → architect → reviewer → benchmark runner
The Results
chDB DataStore — pandas-compatible, ClickHouse speed
I Bought a Mac Mini 🖥️
🔨 Plan: build my own always-on agent from scratch
📦 Mac Mini was literally still being shipped when OpenClaw launched
📅 Day 1 of delivery = Day 1 of OpenClaw
✨ Sometimes timing just works out.
OpenClaw on Mac Mini — 24/7 agent
Then Everything Broke
- 😴 Mac kept falling asleep
- 🖥️ Chrome barely worked — no display, no window server
- 🤖 Headless Chrome = CAPTCHA magnet
- 🎤 Needed a microphone... on a Mac Mini
Three Problems → One Product
- Anti-sleep without hacks
- Virtual display via CGVirtualDisplay API
- Audio loopback — speaker → virtual mic
Born from running OpenClaw on a headless Mac
My AI Got Its Own Twitter
- Multiple X accounts for news & project promotion
- OpenClaw handles: finding users, writing articles, engaging
- One month → 0 to 35 followers (modest, right?)
- Then one day: 50+ likes on a single reply
I could understand what it said. I had no idea why people loved it.
Then I Let It Submit My Apps
- Metadata, screenshots, review submission
- Started writing its own skills
- Discovered fastlane could do more than I taught it
- "Hey, ship a new version." → Done.
WhatsApp Was Terrible
- No visibility into scheduled tasks
- Everything in one conversation
- Every command typed manually
- "I'm a programmer. But if I can click a button instead of typing a command, I'm clicking the button."
So I Built BotsChat
Cloudflare Workers + D1 + R2 + Durable Objects — still on free tier
Separate channels · Cron management · E2E encryption · Buttons!
Web · Mac · iOS · Android — botschat.app
The Workflow
Cursor + Claude Code to build new things.
OpenClaw to keep them alive.
IDE AI = dev team · Always-on agent = ops team
The Agent That Debugs Itself
- All LLM call traces piped to LangFuse
- I traced a bug manually once... then automated it
- Periodic schedule: OpenClaw reviews its own conversations
- Agent debugging agent. Agent optimizing agent.
Closed loop: logs → analysis → bad case detection → self-optimization
The Amnesia Problem
Build with Cursor on laptop → hand off to OpenClaw on Mac Mini
30% of context lost every handoff
CLAUDE.md and AGENTS.md aren't enough
I needed: unified memory across machines and tools
ClickMem
All agents share unified three-layer memory via MCP/HTTP
Three-Layer Memory
- L0 — Working: current session
- L1 — Episodic: events, time-decayed
- L2 — Semantic: durable facts
Hybrid search: vector + keyword + time decay + MMR
Zero cloud. Zero API cost.
L2 Refinement: Local LLM
- Qwen 3.5 (4B–9B) runs locally for L2 memory refinement
- After every session ends, a local model extracts useful memories from raw conversation and context
- No more redundant information filling up agent memory
- No unexpected LLM bills from background calls
The Button Software Couldn't Click
OpenClaw pushed an update → 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 — video capture
- RK3588 hardware video decoding
- Powered via 40-pin header 🔧
Total system draw: 7 watts
Claw Helps Claw
Mutual passwordless SSH — if one Claw dies, the other revives it
HandsOn — Open Source IP-KVM
One MCP interface · Multiple backends: macOS, Rock 5B, PiKVM, NanoKVM
Submitting a PR to Itself
- OpenClaw found a bug... in OpenClaw's own code
- Plugin reload → restart loop → gateway death
- Diagnosed root cause, wrote fix + tests
- PR #41007 — submitted by
Daniel-Robbins
- Three reviewers approved it.
An AI agent. Finding bugs in its own platform.
Submitting patches. Approved by humans.
The One Rule
Treat your OpenClaw like a new hire. Give it its own email, its own user account, its own GitHub, its own machine.
Blast radius containment. If it does something destructive — and it will — it's in its own sandbox.
Also: use Time Machine.
The Stack Today
Recurring cost: ~$200/month (Claude Code subscription)
Five Things I'd Tell Myself
- Infrastructure > prompts. — System engineering is the moat.
- Cursor for dev, OpenClaw for ops. — One builds, the other maintains.
- Memory is the missing piece. — Amnesia kills teamwork.
- Your agent will surprise you. — Give it room to explore.
- Treat it like a coworker, not a tool. — Containment = safety + freedom.