Original video in Chinese.
Key Takeaways
- newtype OS is my content-focused multi-agent system, with eight agents and several skills. It accounted for about 60% of my AI use at the time of this demonstration.
- Install it with npm, connect a provider with /connect, and assign suitable models to agents with /agent-models to balance cost and capability.
- I run its terminal interface in Zed, initialize project knowledge with /init-deep and preferences with /init-soul, then talk to Chief. The system also supports WeChat and invocation by external agents.
Here is a complete introduction to installing and using newtype OS.
I built it this year for people who use AI to digest knowledge and produce content, as I previously did with Claude Code. Because it is customized for that use, I find it more effective for my content work.
At the time of this account, I used newtype OS for around 60% of my AI time, Perplexity for 30%, and Grok and Gemini for the remaining 10%. It had helped me produce tens of thousands of words. If you find it useful, please give the project a star on GitHub.
Install and start it
Run this command in a terminal:
npm install -g @newtype-os/cliInstallation should take only a short time. If it fails, check whether Node.js is installed and obtain it from the official website if needed.
Then enter:
ntThose two letters, short for newtype, start the application.
Connect a model provider
The first configuration step is connecting a provider. You can use an API key and pay for usage, or use a subscription.
Enter /connect to see the provider list. It included major options such as OpenAI, Zhipu, and MiniMax.
I used GitHub Copilot. Although its context window had limitations, it gave me access to major models with little hassle. Select it and follow the account-linking instructions.
Assign models to agents
The system’s foundation is multi-agent orchestration. I designed eight agents for content creation, each with a different role. Not every agent needs the strongest model; doing so may cost more and run more slowly.
Enter /agent-models to see the agents and their current models. Choose an agent to view the models available from your connected provider.
Because I had connected GitHub Copilot, I had a wide selection. After configuring a model for each agent, the system was ready.
Choose a working environment
newtype OS runs in a terminal. You can use a system terminal or an editor such as VS Code. I chose Zed for three reasons.
First, it rendered the terminal UI correctly. I had encountered sluggish or incomplete rendering in VS Code.
Second, I wanted to browse documents and make manual edits while the system ran. I do not ask AI to do everything.
Third, I liked how it looked. The interface in the video is newtype OS running inside Zed, which was free to use.
Initialize the repository and preferences
After opening your content repository, I recommend two built-in commands.
/init-deep asks AI to read the repository’s documents, understand the material, and create KNOWLEDGE.md with the project’s basic context.
/init-soul creates SOUL.md, defining Chief’s personality. I designed two layers: a core personality with behavior rules users cannot change, and a surface personality in SOUL.md that users can edit to match their preferences.
These commands initialize the system’s understanding of the project and the user.
Work through Chief
For daily use, you do not need to decide which agent to invoke. Talk to Chief. It interprets and decomposes the request, then assigns work to the appropriate agents.
I also built in skills for particular capabilities. For deep analysis, Chief could automatically use Super Analyst. Several MCP integrations were included as well.
Connect WeChat or another agent
I integrated idoubi’s WeClaw project. Run:
nt wechat setupScan the QR code in WeChat, then run:
nt wechat startI also added commands so that agents such as OpenClaw could invoke newtype OS as a specialist content team.
For that integration, try:
nt initIt adds the CLI usage instructions to your agent as a skill.
The GitHub README covers these operations. If anything remains unclear, consult it or bring questions and suggestions to the community.
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.