With all the OpenClaw hype this week (Forbes, CNET, IBM, the name changes), I figured people might want a real user report instead of more hot takes. I've been running it for 7 days on a Mac Mini M4 managing two actual businesses.
Setup: Mac Mini M4 16GB from Costco (~$640 after CA tax), running 24/7. Claude Opus 4.5 via Claude Max. Connected to file system, terminal, web browser, messaging, project management, CRM, and GitHub CLI. Use case: autonomous AI employee ("Nigel") for a SaaS product and an AV rental business. Cron-based check-ins, reads/writes its own memory files, runs overnight work shifts while I sleep.
What Actually Works:
Overnight autonomous work is the killer feature. Directive before bed, wake up to structured deliverables in organized directories. Research reports, competitor analysis, lead lists, content drafts - all genuinely usable, not demos.
API integrations. It figured out OAuth flows, diagnosed auth issues, and connected to 4 new services autonomously in one session. This is where it starts feeling like a real employee.
The stuff you'd never get to. It found an SMS chatbot that had been broken for 10 months (legacy app version issue in my automation platform). Diagnosed the root cause, upgraded the components, rewrote the bot prompt through 6 iterations by analyzing real customer conversations. I never would have gotten around to that.
Memory actually works. Daily markdown logs + curated long-term memory + semantic search via embeddings. It remembers multi-day project context, references past decisions, and builds on previous work. Not perfect, but way beyond starting fresh every conversation.
What Doesn't Work:
It hallucinated about itself. When writing marketing copy, it invented $150 in freelancing revenue that never happened and got the hardware price breakdown wrong. An AI making up facts about itself is ironic and would've been embarrassing on a public forum. Everything needs review.
Browser automation is brittle. Rich text editors, complex multi-step UIs - still breaks often enough that you can't fully trust it.
Too agreeable. Without explicit "push back on bad ideas" rules in its personality config, it just says yes to everything. Useless when you need honest feedback. Had to build that behavior in manually.
The Numbers (Honest): Revenue generated: $0. Setup time: ~3 hours initial, then ongoing tweaks. Overnight deliverables: 8 major reports/analyses in one night. APIs connected autonomously: 4. Bug fixes filed: 8 (in a PR I still need to review). The SMS bot fix alone would've taken me a full day I didn't have.
Would I Recommend It? If you're a solo founder running actual businesses and willing to invest setup time: yes. The overnight autonomous work alone justifies the Mac Mini cost. But set expectations right - it's a very capable junior employee, not a senior one. You still need to review everything, catch hallucinations, and handle anything requiring judgment.
If you just want a chatbot: this is massive overkill. Use the API directly.
Happy to answer questions about the setup, what worked, what didn't, or anything else.
With all the OpenClaw hype this week (Forbes, CNET, IBM, the name changes), I figured people might want a real user report instead of more hot takes. I've been running it for 7 days on a Mac Mini M4 managing two actual businesses.
Setup: Mac Mini M4 16GB from Costco (~$640 after CA tax), running 24/7. Claude Opus 4.5 via Claude Max. Connected to file system, terminal, web browser, messaging, project management, CRM, and GitHub CLI. Use case: autonomous AI employee ("Nigel") for a SaaS product and an AV rental business. Cron-based check-ins, reads/writes its own memory files, runs overnight work shifts while I sleep.
What Actually Works:
Overnight autonomous work is the killer feature. Directive before bed, wake up to structured deliverables in organized directories. Research reports, competitor analysis, lead lists, content drafts - all genuinely usable, not demos.
API integrations. It figured out OAuth flows, diagnosed auth issues, and connected to 4 new services autonomously in one session. This is where it starts feeling like a real employee.
The stuff you'd never get to. It found an SMS chatbot that had been broken for 10 months (legacy app version issue in my automation platform). Diagnosed the root cause, upgraded the components, rewrote the bot prompt through 6 iterations by analyzing real customer conversations. I never would have gotten around to that.
Memory actually works. Daily markdown logs + curated long-term memory + semantic search via embeddings. It remembers multi-day project context, references past decisions, and builds on previous work. Not perfect, but way beyond starting fresh every conversation.
What Doesn't Work:
It hallucinated about itself. When writing marketing copy, it invented $150 in freelancing revenue that never happened and got the hardware price breakdown wrong. An AI making up facts about itself is ironic and would've been embarrassing on a public forum. Everything needs review.
Browser automation is brittle. Rich text editors, complex multi-step UIs - still breaks often enough that you can't fully trust it.
Too agreeable. Without explicit "push back on bad ideas" rules in its personality config, it just says yes to everything. Useless when you need honest feedback. Had to build that behavior in manually.
The Numbers (Honest): Revenue generated: $0. Setup time: ~3 hours initial, then ongoing tweaks. Overnight deliverables: 8 major reports/analyses in one night. APIs connected autonomously: 4. Bug fixes filed: 8 (in a PR I still need to review). The SMS bot fix alone would've taken me a full day I didn't have.
Would I Recommend It? If you're a solo founder running actual businesses and willing to invest setup time: yes. The overnight autonomous work alone justifies the Mac Mini cost. But set expectations right - it's a very capable junior employee, not a senior one. You still need to review everything, catch hallucinations, and handle anything requiring judgment.
If you just want a chatbot: this is massive overkill. Use the API directly.
Happy to answer questions about the setup, what worked, what didn't, or anything else.
How do you defend against prompt injection?