AI Agent · MCP · Context Engineering
I build agent-native products, developer workflows, and digital gardens for "farming in the cyber world".
I'm 晓灰, a "cyber farmer" who writes code and words. My work sits at the intersection of AI agents, software engineering, knowledge workflows, and product systems.
- Agent expert at Ant Group, working on production agent systems for insurance scenarios.
- Lead agent work for a million-MAU insurance product, including product interpretation, quality scoring, and evaluation pipelines.
- Creator of MCPAdvisor, a tool for discovering and installing MCP servers.
- Built
writingHelper, which has served 15,000+ users. - PKU '21. Based in Beijing.
- Agent-native products: systems where agents close the loop between user intent, production workflows, evaluation, and product improvement.
- MCP and tool ecosystems: practical ways to select, route, and compose external capabilities for agents.
- Context engineering: memory, retrieval, state, event streams, and runtime patterns that make agents more reliable.
- Digital gardens and writing systems: tools and habits for turning scattered notes into public, useful knowledge.
| Project | What it does | Stack |
|---|---|---|
| MCPAdvisor | A discovery and installation assistant for MCP servers, built around the problem of choosing the right tool for a specific workflow. | TypeScript, MCP |
| writingHelper | Grammarly and Linggle-like writing assistance in one workflow, used by 15,000+ people. | HTML, browser tooling |
| ai-speeds | A simple proxy that helps Claude Code work with OpenRouter or DeepSeek. | TypeScript |
| cc4pm | An agent harness optimization system for Claude Code: skills, memory, instincts, safety, and research-first development. | JavaScript, agents |
I write at xiaohui.cool, a public digital garden about AI agents, MCP, software engineering, learning, and social observation.
- Agent Native: letting products form intelligent loops
- From Claude Code to OneAgent: context engineering in practice
- How to build reliable agent systems
- OneAgent + MCPs: creating domain agents quickly
- Farming in the Cyber World
- Building production-grade agent runtimes, evaluation loops, and product event streams.
- Turning rough AI workflows into reliable, observable, and reusable systems.
- Writing about the interfaces, infrastructure, and habits behind agent-native productivity.
- Blog: xiaohui.cool
- Resume: istarwyh.github.io/resume-it
- GitHub: github.com/istarwyh