AI Agent developer. I build local-first AI agents β agents that actually do the work: break down tasks, call tools, stream every step live, remember what they learned, and turn repeat work into automation. All running on your machine, owned by you.
Years of production engineering (JVM microservices β Rust systems) went into this stack β and these days I work with agents as much as on them: the ecosystem below is designed, built, reviewed, and shipped in tight loops with my own agents. Bodhi helps build Bodhi.
An end-to-end, open-source desktop AI agent I design and build from the runtime up:
| Project | What it is |
|---|---|
| π§ Bodhi-AI | The desktop agent app β a Tauri shell over the Bamboo runtime + Lotus UI. Watch the agent think, act, and finish. |
| π Bamboo-agent | The heart: a local-first AI agent runtime in Rust. Persistent memory, 20+ built-in tools, skills, MCP, sub-agents, workflows & schedules behind one HTTP + WebSocket API. Embed it as a crate or run it as a server. |
| πͺ· Lotus / lotus-next | The living interface for agents β React UI that streams reasoning, tool calls, and task lists live, turning a black-box AI into a transparent teammate you can watch and interrupt. lotus-next is the mobile-first shadcn rebuild. |
| β¨ Nova | Give an LLM control of the macOS desktop: a single self-contained Rust binary speaking MCP β screenshots, Set-of-Mark targeting, Apple Vision OCR, mouse & keyboard. |
| π Zenith | The monorepo that ties it all together β runtime, UI, shell, Go backend, docs, and a coordinated release train. |
| π― Pavilion | Bodhi AI's front door β bilingual (δΈζ/EN) website & docs with long-form architecture deep-dives. |
- Local-first & private β your model keys, your data, your machine; the cloud is optional
- Transparency over magic β every token, tool call, and decision streamed live to the UI
- Memory that compounds β persistent, searchable agent memory (CJK-aware BM25 + consolidation), not goldfish sessions
- Real work, not chat β task decomposition, sub-agents, schedules, workflows, and desktop control via MCP
- Agent-native development β multi-agent implementation sweeps, adversarial review passes, scheduled autonomous runs: daily practice, not demos
- Rust where it counts β one fast, embeddable runtime instead of a stack of glue
Daily drivers: Rust Β· TypeScript/React Β· Tauri Β· MCP Β· WebSocket streaming Also fluent in: Java Β· Kotlin Β· Scala Β· Go Β· Vue.js
Feel free to reach out at mugeng.du@gmail.com.
