Original video in Chinese.
Key Takeaways
- My local OpenCode setup with newtype Profile handles creation and knowledge management; Kimi Claw handles cloud collection and automation. Email brings the results into local Obsidian files and avoids relying on the home computer staying connected.
- Kimi Claw is a hosted OpenClaw setup driven by K2.5. In the version discussed here, deployment was simple, it supported up to eight parallel subagents, and skills could be added through Claw Hub.
- I preferred this hosted option to buying a Mac mini solely for OpenClaw, based on my needs for isolation, cost, and convenience. At the time, it was a beta available to members.
I had been out of conventional employment for almost a year, but recently installed Feishu again just to use Kimi Claw.
This hosted version of OpenClaw completed a missing part of my AI operating system.
Think of my local AI OS as a home base that must stay safe and stable, and Kimi Claw’s cloud virtual machine as an external production chain. It can handle demanding work elsewhere, as long as the results come back.
Why I chose a hosted setup
OpenClaw was the first personal agent system to gain global attention on that scale and, in my view, the second phenomenon after ChatGPT. It showed many people what AI working through a terminal could do.
I had not made a video about it because I disagreed with buying a Mac mini solely to run it.
Such a machine would generally still call models through APIs rather than run the large model itself. And if you bought a separate computer because you feared privacy leaks or accidental deletion on your main machine, a cloud deployment was worth considering.
Earlier hosted options had two problems in my experience. Setup was difficult: I spent an hour trying a Cloudflare deployment and gave up. Also, agent systems consume far more tokens than a chatbot, so capable and economical models were essential.
I eventually chose Kimi Claw.
Deployment and messaging
I liked its simple deployment. At kimi.com/bot, clicking a button configured compute and 40GB of storage in about a minute.
To use it through Feishu, I created an application on Feishu’s developer platform, added bot capability, configured permissions and message-receiving events, supplied the App ID and Secret to the integration, and published the application. The logic was similar to connecting Discord.
Alternatively, you could use the Kimi app directly. At the time of the episode, the international app had updated while the Chinese app was still under review. The Kimi Claw entry appeared above the input field.
Although the interface still looked conversational, the system behaved differently from ordinary Kimi. It analyzed tasks, found approaches, and could run up to eight subagents in parallel in the background.
K2.5 made this possible. I had previously introduced its Agent Swarm capability in a separate video.
In my experience with OpenClaw using some overseas model APIs, multiple subagents could trigger Too Many Requests errors. My Kimi Claw use did not encounter that issue. I also considered K2.5’s pricing suitable for the substantial token usage of multi-agent systems.
Add capabilities through Claw Hub
Kimi Claw can be customized. Ask it to search Claw Hub for a skill and install it.
I described Claw Hub as an app store for agents, with more than 5,000 skills at that time, spanning development and office work. For example, I had it install a Xiaohongshu skill so it could write in that platform’s characteristic style.
Bring scheduled results into Obsidian
My more advanced use was scheduled information collection, such as gathering AI news every eight hours.
The results stayed in the cloud. Copying them into local files by hand would have been inconvenient, so I used email as a bridge.
First, I installed a lightweight email-sending tool in Kimi Claw and configured it with my QQ Mail SMTP password so it could send new results.
Second, I installed an email-receiving plugin in Obsidian. It supplied an address whose incoming messages were saved automatically as notes.
Finally, I gave that destination address to Kimi Claw and asked it to email the results of scheduled tasks. An Obsidian note appeared with the collected AI news.
This preserved the cloud machine’s separation from the local environment while automatically bringing the data I needed into local files.
The complete division of work
Local OpenCode handled deep creation and knowledge management. Cloud Kimi Claw handled information collection and automation. Email connected the two.
I consider OpenClaw an important product. Given my privacy, safety, and cost concerns, I preferred a hosted option such as Kimi Claw with K2.5 for long-term use.
At the time of this episode, Kimi Claw was still a beta initially open to members. For ordinary users, I thought it better value than buying separate hardware and then paying for inference as well.
That is all for this episode. If you want to understand AI, become a more capable independent creator, and meet people with similar interests, join the newtype community. See you next time.