My background is in networking, security, and operations. I did not begin as a software developer. AI coding tools changed that boundary for me: they made it possible to turn years of practical experience, stubborn product opinions, and a lot of testing into working open-source software.

I am interested in AI agents, computer use, local deployment, Chinese speech, memory, and the unglamorous product work around them. Installation, permissions, updates, logs, failure recovery, and honest evidence matter as much as a good model response.

I write about the experiments on KevinAIStack, my WeChat public account. Earlier projects taught me a great deal about Hermes, OpenClaw-style personal assistants, memory systems, local speech, and running agents on a Mac mini. Penglai is where those lessons became a product of my own.

Stand on the shoulders of good open source, then add the judgment that comes from using it yourself.

• Local first. Your data, tools, and project history should remain understandable and controllable.
• Harness matters. A model can be brilliant, but the surrounding permissions, tools, memory, and recovery path decide whether it is useful.
• Evidence before claims. Source tests, packaged tests, installed tests, and real external-account tests are different things.
• Lower the entrance fee. People should not need to become terminal experts before an agent can help them.
• Build one honest step at a time. A smaller thing that works every day is worth more than a large promise.

Penglai 0.5.5 packages the official DeepSeek Harness into a desktop application for Apple Silicon, Intel Mac, and Windows x64.

DSH is the only agent core. Penglai adds the parts that make it usable as a personal computer product: first-run setup, process supervision, updates, uninstall, a signed Plugin Center, Office, Workspace-isolated Memory, Weixin and Feishu messaging, and optional local speech recognition and voice generation.

Office and Memory are included and enabled for a fresh installation. Messaging, SenseVoice, MOSS-TTS, and Companion wait for the user. The source is MIT licensed, the public Release is immutable, and the trust limits are stated plainly.

AI agents · Computer use · Local deployment · Chinese NLP and speech · Memory · Desktop packaging · Security and operations

On KevinAIStack I keep long-form notes about the parts that rarely fit in a launch post: what failed, what changed my mind, which open-source projects were worth learning from, and what it takes to keep a personal AI system useful after the first demo.

Topics have included running large agent workloads, comparing memory systems, moving local speech from Whisper to FunASR, and operating several coding agents on a Mac mini.

_{Search WeChat for KevinAIStack, or scan the code.}

我做过十多年网络、安全和运维,并不是从软件开发者起步。AI 编程工具改变了这条边界: 它让我可以把实际经验、产品判断和大量测试,变成真正运行的开源软件。

我长期关注 AI Agent、Computer Use、本地部署、中文语音和记忆系统。对我来说,模型回答得聪明只是开始。安装是否顺利、权限讲不讲得清、升级会不会把用户卡住、失败以后能不能恢复,同样决定一个 Agent 能不能进入日常生活。

微信公众号 KevinAIStack 记录了这些实践。从 Hermes、OpenClaw 一类个人助手,到记忆系统、本地语音和 Mac mini 上的长期运行,我玩过很多方案。蓬莱是这些经验第一次真正收拢成自己的产品。

站在优秀开源项目的肩膀上,再加上自己长期使用后形成的判断。

• 本地优先。 数据、工具和项目历史应该在用户手里,而且看得懂、管得住。
• 马具很重要。 模型像千里马,权限、工具、记忆和恢复流程决定它究竟能跑多远。
• 证据先于结论。 源码测试、打包测试、安装后测试和真实外部账号测试不能互相冒充。
• 降低使用门槛。 普通人不应该先成为终端专家,才有资格用上 Agent。
• 一步一步做真的。 一个每天能用的小东西,比一个宏大的承诺更有价值。

蓬莱 0.5.5 把官方 DeepSeek Harness 装进桌面客户端,支持 Apple 芯片 Mac、Intel Mac 和 Windows x64。

DSH 是唯一 Agent 核心。蓬莱补上一个个人电脑产品真正需要的部分: 首次引导、进程守护、升级、卸载、签名插件中心、蓬莱办公、按 Workspace 隔离的蓬莱记忆、微信与飞书消息,以及可选的本地语音识别和语音生成。

办公和记忆在全新安装中默认启用。手机消息、SenseVoice、MOSS-TTS 和主动陪伴等待用户自己开启。源码采用 MIT 许可,公开 Release 不可变,做到了什么和没有做到什么都会讲清楚。

KevinAIStack 主要写那些发布文章里经常放不下的内容: 什么失败了,哪些判断被实践推翻了,哪些开源项目值得认真学习,以及一个个人 AI 系统过了演示阶段以后,怎样才能继续有用。

写过的话题包括大规模 Agent 使用、不同记忆系统的真实差异、把本地语音从 Whisper 换到 FunASR,以及在一台 Mac mini 上协作使用多个编程 Agent。