After remotely installing OpenClaw for dozens of people, the three questions I get asked most are: Is it safe? Is it expensive? And which should I pick, Claude Code or OpenClaw? Let's address them all here.
Is OpenClaw Safe?
I wrote about safety in a dedicated post before; here are a few updates.
OpenClaw now ships a new version every day. Version 3.8 came out the day before yesterday. The permission-overreach and Feishu (Lark) pairing issues discussed by the community earlier have all been fixed in this release. Default permissions are now set conservatively—you can configure it so it can't delete or modify files, and the permission docs are pretty clear.
For example, in February someone installed an older version where Feishu private chat was open by default—if you created a bot in your organization, anyone could click in and talk to it, which meant they could control your computer. That was genuinely dangerous. Now the Feishu plugin requires pairing approval by default; the first time must be manually authorized on your machine.
But safety ultimately depends on your assets. Attacking you costs effort; the attacker has opportunity costs and risks getting caught. If what's stolen doesn't cover those costs, no one will target you.
So first ask yourself: what's actually on your computer? If it's just personal daily files, don't be too anxious. But do consider: are you using a company machine? Your personal stuff may not be valuable, but company information matters to the company—that's something to think about. Take it seriously when needed: buy a separate machine for physical isolation, pick a better model. The heavier the assets, the more you should spend on protection—it's the same logic as buying a security door.
Staying in sync with official updates matters more than anything else.
Is OpenClaw Expensive to Use?
I see a lot of "it's too expensive" takes. I'm not always sure which kind of expensive people mean.
If you're talking about top-tier overseas models like Claude Opus or the GPT series, direct API calls are indeed not cheap. The models themselves are pricey—overseas users routinely spend hundreds of dollars, and those posts scare a lot of people.
The domestic situation is completely different. Domestic models generally follow a cost-performance strategy. DeepSeek famously slashed prices to 5% of OpenAI's rates—that was its biggest impact at the time, the "price butcher." Now there are tons of packages on the market with several times the volume at 1/5 or even 1/7 the overseas price—a gap of nearly 20x.
Here's a concrete example. MiniMax has a ¥200/month plan that refreshes quota every 5 hours; one person couldn't use it all up.
What does ¥200 mean? Tencent Video is a few dozen per month, two cups of coffee run over a hundred. Entry level is even cheaper—some platforms offer first-month discounts under ¥10 to get started, and normal basic usage is around ¥30–40 a month.
If you think even ¥40 is too expensive, I think that's a shame. Alibaba ran a promotion recently where they gave you a cup of milk tea—drinking it probably changes nothing. Spending the same money on AI, at this point in 2026, might genuinely change your understanding of AI entirely.
Claude Code or OpenClaw?
I'm pretty firm on this one.
If you have the means, use Claude Code. The "means" are two things, and honestly they're somewhat demanding: first, you need to be able to access the overseas internet; second, you need to be able to subscribe on the Claude or ChatGPT official website, which usually requires an overseas friend to help you with a credit card. These two barriers stop a lot of people.
If you can clear them, I strongly recommend trying it. You'll feel a 10x level of productivity change—it's no exaggeration. When I use Claude Opus 4.6, I often feel—wait, it just finished this by itself? The more ordinary you are, the bigger the impact might be.
If you can't access the overseas internet or figure out overseas payment, just install OpenClaw. Domestic models already perform quite well for coding scenarios; the gap isn't huge. The advantage of OpenClaw is simplicity—install and chat, everyone knows how to chat, so it fits easily into your daily routine. I wrote about the difference between the two in another post: OpenClaw is more like a Manager, while Claude Code is a Worker. A Manager only has value when the Worker's execution is solid.
And OpenClaw is evolving fast. Many skills will be coming online soon, and it will connect to more and more tools. Overseas model providers are also accelerating their own ecosystems. Both sides are running; the gap will only get smaller.
About Remote Installation
One more thing. Some people ask whether my free remote installation is safe—does opening a tunnel mean I can control your computer?
Essentially, yes. But I can only operate while you actively have the tunnel open; close it after installation and you can uninstall everything related to the tunnel.
The entire installation process is done by an AI agent, not me typing commands manually. After dozens of installations, this agent is very well-behaved—it finishes the task and shuts down cleanly, doesn't mess around. If you're worried, I can share my screen with you and you can watch the AI operate the whole time. It's actually pretty interesting; you might get a stronger sense of how far AI has come. That's also why I install it for free—every installation helps the AI agent evolve.
Nothing is 100% safe, and this certainly isn't. But I genuinely have no reason to mess with your computer.