Ted Huang · Open sourceTed Huang · 開源作品

Working software, built in the open.

能用的軟體,公開打造。

18 public projects from an engineer with twenty years in IT and security who now builds AI agent tooling. Each one has its own page in English and Traditional Chinese, and every page says what doesn't work yet.

18 個公開專案,來自一個做了二十年 IT 與資安、現在專心打造 AI agent 工具的工程師。每個專案都有自己的中英雙語介紹頁,也寫明還做不到的地方。

EN / 繁中Honest status on every page每頁都寫明目前狀態Source on GitHub原始碼都在 GitHub
Project directory專案目錄18 projects18 個專案
01AI toolsAI 工具Multi-model chat, review, memory多模型對話、審查、記憶6
02Agent infrastructureAgent 基礎建設Connectors, guardrails, memory連接器、護欄、記憶6
03Security and IT資安與 ITScanner, control plane, training掃描、控制平面、教材3
04Small things and notes小作品與筆記Toy, notes, utility玩具、筆記、工具倉庫3
18public projects個公開專案
20years in enterprise IT and security年企業 IT 與資安經驗
2languages on every page種語言,每一頁都有

From MUD wizard to AI agents

從 MUD 巫師到 agent

Code changes. The ideas don't. The heartbeat I learned about in a MUD is what keeps my agents running today.

Code 怎麼變,觀念都不會變。
我在 MUD 學到的心跳,現在用在 agent 上。

  1. Text MUD

    I got hooked on a text MUD as a kid and was still playing it in college.

  2. Wizard and sysadmin

    While I studied computer science at National Central University in Taiwan, I went from player to wizard, coding for that world and keeping its server running.

  3. Corporate IT

    Then I spent nearly twenty years at US companies, working on identity, endpoints, ERP, BI, CRM, and security.

  4. Agents with a heartbeat

    Long-running agents need a heartbeat too, a steady tick that keeps the work moving.

  • from player to wizard
  • then corporate IT
  • agents need a heartbeat too
I build AI agents with habits I picked up in the MUD and in twenty years of corporate IT. A MUD runs on heart_beat: every mob and NPC has one, and if it stops, poison never kicks in and nobody heals. Long-running agents need the same heartbeat.
  1. 文字 MUD

    我從小迷上一個文字 MUD,一路玩到大學。

  2. 大學當巫師

    在中央大學念資工時,我從玩家變成巫師,
    幫那個世界寫程式,也顧它的主機。

  3. 企業系統

    之後在美國將近二十年,
    在企業裡做身分管理、端點、
    ERP、BI、CRM 和資安。

  4. 同樣的心跳

    長時間跑的 agent 也要有心跳,
    事情才會一直往前走。

  • 玩 MUD,後來當巫師
  • 之後做企業系統
  • agent 也要有心跳
我小時候迷上一個文字 MUD,一路玩到大學。在中央大學念資工時,我從玩家變成巫師,幫那個世界寫程式,也顧它的主機。之後在美國將近二十年,在企業裡做身分管理、端點、ERP、BI、CRM 和資安。現在我用同樣的習慣做 AI agent 工具。heart_beat 是 MUD 最基本的概念,每一個 mob、每一個 NPC 都有心跳,世界靠心跳在流動。心跳沒在跑,毒不會發作,也不會回血。
長時間跑的 agent,也需要同樣的心跳。

01 · AI tools01 · AI 工具

AI that works where you already work.

在你熟悉的工具裡用 AI。

Ask ChatGPT, Claude, Gemini, and Grok together and let them check each other, from the desktop, the browser, the terminal, or Claude Code. Plus a companion that remembers your day on your own machine.

讓 ChatGPT、Claude、Gemini、Grok 一起回答、互相檢查,桌面、瀏覽器、終端機或 Claude Code 裡都能用。另外還有一個在你自己電腦上記住一整天的陪伴 app。

A local workbench that runs staged multi-agent coding workflows on the headless runtimes of Claude Code, Codex, Grok, and Antigravity, with gates and a replayable evidence feed.

本機工作台,在 Claude Code、Codex、Grok、Antigravity 的 headless runtime 上跑分階段的多 agent 工作流程,有關卡,也有可重播的證據紀錄。

TypeScript · React · Node.js · Tauri 2

AI-SisterAI-Sister

Active持續開發

A local-first desktop companion that remembers what you worked on, stays quiet most of the time, and cites a source for every answer.

本機優先的桌面陪伴:記得你做過什麼,大部分時間保持安靜,每句回答都附上出處。

Tauri 2 · Rust · SQLite · Windows OCR

A small web app for idea review: pick 5, 12, or 16 rounds, leave, and come back to a Markdown report by link or email.

點子審查的小型網頁 app:選 5、12 或 16 輪,先離開,之後用連結或 email 拿到 Markdown 報告。

Hono · React · SQLite · OpenRouter

Picture a human employee

想像一個人類員工

I start with how a real employee would do the job, and only wake the AI where that person would stop and think.

先想一個真人員工會怎麼做,
只有他得動腦的地方才叫醒 AI。

  1. Open the ERP

    An employee opens the ERP to write up a quote.

  2. Plain code

    Anything that takes no thinking runs as code: CRUD, search, notifications.

  3. Wake the AI

    The AI only wakes up for a judgment call, like how big a discount this customer gets.

  4. Finished in the database

    In a 10-step process, maybe 8 steps finish there and only 2 wake the AI.

  • routine steps run as code
  • judgment calls wake the AI
  • maybe 8 of 10 steps are just code
Agents should work like that employee: analysis, judgment, and writing go to the model, and everything else is plain code. Designing one comes down to two questions: if this were a company, how would it manage its people, and if it were a person, how would they do the job? Then I turn the answers into code.
  1. 打開 ERP

    員工打開 ERP,要開一張報價單。

  2. 走 code

    不用動腦的事走 code,
    像 CRUD、搜尋、通知。

  3. 叫醒 AI

    要動腦的才叫醒 AI,
    像是這個客戶該給多少折扣。

  4. 資料庫做完

    10 步的流程,可能 8 步在資料庫做完,
    只有 2 步要叫醒 AI。

  • 例行步驟走 code
  • 要判斷才叫醒 AI
  • 10 步可能 8 步走 code
想像一個人類員工。他不會每次打開 ERP 都要想怎麼建報價單,點幾個按鈕就好了,只有在判斷這個客戶該給多少折扣的時候,才需要動腦。Agent 也一樣:動腦的事給 LLM,像分析、判斷、寫作;不動腦的事走 code,像 CRUD、搜尋、通知。一個 10 步的流程,可能 8 步在資料庫完成,只有 2 步需要叫醒 AI。AI Agent 的設計,最終就是在回答兩個問題:
如果這是一間公司,它會怎麼管理員工?如果這是一個人,他會怎麼做這件事?然後把答案寫成 code。

02 · Agent infrastructure02 · Agent 基礎建設

Plumbing for agents that run all day.

Agent 整天跑,也不會出事。

Connectors, guardrails, memory, and usage tracking for coding agents such as Claude Code, Codex, and OpenClaw.

給 Claude Code、Codex、OpenClaw 這類 coding agent 用的連接器、護欄、記憶與用量追蹤。

An unofficial connector that lets AI agents watch, nudge, and orchestrate Better Agent Terminal sessions, with opt-in tiers and deterministic gates.

非官方連接器,讓 AI agent 查看、推動與調度 Better Agent Terminal 的 session,權限分層、需要主動開啟,並有確定性的關卡。

Python · MCP · CLI

bat-coworkbat-cowork

Early stage早期階段

A server-first cowork workstation: a Paseo-derived daemon with a Better Agent Terminal desktop client. Work in progress, not a release.

伺服器優先的協作工作站:Paseo 衍生的 daemon,加上 Better Agent Terminal 桌面端。仍在開發中,尚未發布。

TypeScript · Node.js · Tauri

TokenMonsterTokenMonster

Paused暫停開發

Tracks Claude Code, Codex, Gemini CLI, and Grok Build token use on your own machine, with a live dashboard and companions that grow with real milestones.

在你自己的電腦上追蹤 Claude Code、Codex、Gemini CLI、Grok Build 的 token 用量,有即時儀表板,陪伴角色會隨真實里程碑成長。

TypeScript · Electron · Node.js · SQLite

Clawd-LobsterClawd-Lobster

Paused暫停開發

A curated operating layer for Claude Code: reviewed specs, persistent memory, reusable skills, and workspaces that follow you across machines.

Claude Code 的精選運作層:經過審閱的規格、持久記憶、可重用的 skill,以及跨機器延續的工作區。

Python · MCP · SQLite · Claude Code

An MCP server for long-term agent memory: decisions, resolved issues, open questions, and knowledge that you can inspect.

給 agent 用的長期記憶 MCP 伺服器:決策、已解決的問題、待回答的問題與知識,都可以檢視。

Python · FastMCP · SQLite

Done is better than perfect

完成勝過完美

The main project gets to its first milestone before anything else. New ideas wait their turn, and optimizing comes after it's done.

主專案先做到第一個里程碑,
新想法先排隊,做完再來優化。

  1. Parking Lot

    New ideas get written down here. No building, no research, no optimizing.

  2. One main project

    Only one at a time, and I stay with it through the boring parts.

  3. First milestone

    No new projects until it gets there.

  4. Then optimize

    Only finished work gets optimized.

  • one main project at a time
  • new ideas stay parked
  • optimize only what's finished
Until the main project reaches its first milestone, new ideas go into the Parking Lot and stay there. I accept the boring part of execution, and I only optimize what's finished. Every project page here also says what doesn't work yet.
  1. 想法停車場

    新想法先寫進 Parking Lot,
    不做、不研究、不優化。

  2. 一個主專案

    一次只做一個,
    執行期的無聊也接受。

  3. 第一個里程碑

    做到之前,不開新坑。

  4. 再來優化

    完成,才有資格優化。

  • 一次一個主專案
  • 新想法先停著
  • 做完才有資格優化
我一次只做一個主專案。在它做到第一個里程碑之前,新想法只能寫進 Parking Lot,不實作、不研究、不優化。執行期的無聊,我接受。
完成比完美重要,做完了,才有資格優化。
這裡每個專案頁也都寫了還做不到的地方。

03 · Security and IT03 · 資安與 IT

Twenty years of IT, turned into tools.

二十年 IT 經驗,做成工具。

Security checks, machine control, and awareness training that come from running real company systems.

從實際維運公司系統累積出來的資安檢查、機器管控與資安意識教材。

04 · Small things and notes04 · 小作品與筆記

Not everything needs to be a platform.

不是每樣東西都要做成平台。

A parody toy, a strategy notebook, and a helper repository for a community bot.

一個惡搞小玩具、一本策略筆記,還有一個替社群機器人存圖的倉庫。

About關於

An operator who builds.

維運出身,自己動手做。

I studied computer science at National Central University in Taiwan, then spent nearly two decades in the United States running identity, endpoints, ERP, BI, CRM, and security programs inside real businesses.

我在台灣讀中央大學資工,之後在美國將近二十年,在真實的企業裡負責身分管理、端點、ERP、BI、CRM 與資安計畫。

I build AI agent tooling with the same habits: people keep the final say, state survives a restart, and every claim leaves evidence.

現在我用同樣的習慣做 AI agent 工具:最後的決定權留在人手上,狀態重開也不會遺失,每個說法都留下證據。

Working for months in a real environment matters more than looking clever in a demo.
demo 裡看起來聰明,不如在真實環境裡穩定跑上好幾個月。