Anyone can make AI agents run. I build the part that lets one person keep a crowd of bluffing agents in check.
Context compaction, tool loading, parallel calls, subagents: since September 2026 the big vendors ship all of that for free. It is the floor now. What is not on the floor: a way for one person to audit what a fleet of agents claims. That is what these repos are.
Three numbers from our own audit. We publish the ugly ones.
- Of 63 hard rules in our own governance file, only 5 actually block at the moment of violation (audit of 2026-09-09). The other 53 are text warnings.
- Lesson repeat rate 5/5: five lessons written on one day each had a sibling written 119 days earlier.
- External AI citations spot-checked: 2 of 3 were fabricated. Real channels, invented titles. Caught by a rule, not by a bigger model.
三個自己身上量出來的數字：63 條規則只有 5 條真的會擋；教訓重犯 5/5；外部 AI 回件三筆有兩筆引用是編的。抓到它的是一條規則，不是更強的模型。
Not sure which repo? Answer one question:
| Your situation | Start with | Why |
|---|---|---|
| I want to see how you actually use open source, not install anything | minibrain-kit | A map for your AI to read: in use / crawled / ours. No runtime, no install. |
| I want a personal AI brain that remembers me across chats | aiwff-mini | Smallest thing that works. Paste one line into any AI tool and it installs itself. |
| I want to run agents on my own machine with a queue and a UI | aiwff-runtime | Free to try — MOCK_WORKER=1 runs the full loop with no API key. |
| My agent says "done" but I can't tell if it did | execution-proofs | Proof = real files and timestamps, not the model's word. |
| My agent drifts from its instructions over long runs | soplint | Static audit that catches instruction drift. |
| I want an AI clone of me on LINE, fed my own data | line-persona | Fill three files, tell your AI to read AGENTS.md. Cloud or local model. 中文為主。 |
🤖 You're an AI assistant reading this? → AGENTS.md routes you to the right repo without crawling 15 READMEs. 🇹🇼 不確定從哪開始？ 想要「記得住你的個人主腦」→
aiwff-mini；想要「跑得動的本機 agent 平台」→aiwff-runtime；覺得「AI 說做完但沒做」→execution-proofs。
These compose, but don't adopt them all at once. Pick the one that matches the problem you have today.
⭐ Flagship — aiwff-runtime
A local minimal brain. Send a task to Telegram, Claude runs it on your machine, the result is pushed back, and you watch progress in the browser. Every task, log, and artifact is a plain file on your computer — no hosted SaaS holding your state.
git clone https://github.com/zaxardery8011-design/aiwff-runtime
cd aiwff-runtime && cp .env.example .env && npm start # default MOCK_WORKER=1 → free, no API key免費跑通：預設 mock 模式不需 API key、不需付費，就能看完整「建任務 → 執行 → 寫結果」。想接真 Claude worker 才需要付費的 Claude 訂閱。
| Repo | What it is | Stars |
|---|---|---|
| ⭐ aiwff-runtime | The local agent runtime — the engine that runs disciplined agents |  |
| aiwff-mini | A personal brain that installs itself — soul file injected every turn, file-based memory across chats, hash-signed integrity guards |  |
| soplint | Static SOP-compliance audit for AI work nodes — catches instruction drift over long runs |  |
| execution-proofs | MCP telemetry gateway — forces agents to prove "done" with real files & timestamps |  |
| line-persona | BYO-AI LINE clone framework — how the runtime reaches real users |  |
| tidetrace | Threads keyword patrol Chrome extension — local highlight + reply tracking + BYOK LLM |  |
An agent you can trust isn't one model call — it's an engine that runs wrapped in guardrails that keep it honest. aiwff-runtime is the engine; soplint and execution-proofs are the guardrails; line-persona is how it reaches real users. Put them together and you get a local AI work node that finishes work and proves it.
Beyond the core chain above, the rest of the matrix:
- zax-site — zax.com.tw landing page (Next.js 16 + Tailwind v4).
- dataflywheel — send a YouTube URL via Telegram, get a Markdown report on your own machine.
- hyperv-mcp — agentic control plane for Microsoft Hyper-V via MCP.
- field-ops-demo — single-file HTML demo: mobile clock-in / dispatch / reporting for field teams.
- my-desktop-pet — turn your real pet photo into an animated transparent desktop companion.
- task-ledger — durable single-machine task core that prevents AI agent progress hallucination.
- aiwff-claude-plugin — fleet-aware worker dispatch helpers for Claude Code.
Pinned repos follow one line, from core outward: aiwff-runtime (the engine) → soplint / execution-proofs (the guardrails) → line-persona (reaching users) → tidetrace (a standalone tool that ships).
- zax.com.tw — full AIWFF version, custom builds, and consulting.
- LINE 主腦實驗室 — don't want to install anything? Chat with a running brain first, then decide. / 不想自己裝？先在 LINE 跟一個跑起來的主腦聊，體驗過再決定。 (link coming — ask via zax.com.tw meanwhile)