(00) 一個獅群 · 六個 Agent · 2026One pride · six agents · 2026

LionAgents

六個各司其職的 AI Agent:可在你自己的環境執行、自備模型金鑰,重要動作交給人核准,數字只寫實測過的。

Six AI agents, each with one job. They run in your own environment on your own model keys, hand the important decisions to a human, and only quote numbers they have measured.

往下Scroll
  1. 01QueryLion
  2. 02LedgerLion
  3. 03LensLion
  4. 04PulseLion
  5. 05ProbeLion
  6. 06RuleLion

(01) 原則Principles

不是更多自動化,而是可以信任的自動化。Not more automation. Automation you can trust.

下面每一條原則都標出真正做到的成員——不是每個 Agent 都做到每一條。

Each principle lists the members that actually do it — not every agent does everything.

  1. 01

    在你的環境執行Runs where you are

    自架、本機或在你的 CI 裡跑。LedgerLion 資料全在本機 SQLite;RuleLion 加 --no-llm 不發任何網路請求。

    Self-hosted, local or inside your CI. LedgerLion keeps everything in local SQLite; RuleLion with --no-llm makes zero network requests.

    • QueryLion
    • LedgerLion
    • LensLion
    • PulseLion
    • ProbeLion
    • RuleLion
  2. 02

    重要動作交給人Humans approve

    QueryLion 寫入前先看 Dry-run 差異;PulseLion 未核准的草稿永遠不發布;ProbeLion 可在呼叫工具前人工核准。

    QueryLion shows a dry-run diff before any write; PulseLion never publishes an unapproved draft; ProbeLion can pause for approval before each tool call.

    • QueryLion
    • PulseLion
    • ProbeLion
  3. 03

    自備金鑰(BYOK)Bring your own key

    每個 Agent 都能自選模型供應商、貼上金鑰並測試連線;畫面只顯示遮罩。

    Every agent lets you pick the provider, paste a key and test the connection; keys are only ever shown masked.

    • QueryLion
    • LedgerLion
    • LensLion
    • PulseLion
    • ProbeLion
    • RuleLion
  4. 04

    確定性優先,模型其次Deterministic first, model second

    LedgerLion 由規則決定提醒、模型只寫文案;LensLion 先用 axe-core 量測;RuleLion 校驗每條宣稱;PulseLion 的數字全由程式計算。

    LedgerLion’s reminders are fixed rules and the model only writes the wording; LensLion measures with axe-core first; RuleLion verifies every claim; PulseLion computes every number in code.

    • LedgerLion
    • LensLion
    • RuleLion
    • PulseLion
    • QueryLion
  5. 05

    誠實的數字Honest numbers

    實測結果附日期、模型與測試集,連沒達標的項目也寫出來。

    Measurements come with the date, the model and the test set — including the targets they missed.

    • QueryLion
    • LensLion
    • ProbeLion
    • RuleLion

(02) 獅群索引The pride index

六個 Agent,一種語言Six agents, one language

  1. 01/06

    QueryLion

    資料提問Data questions

    可自架的 PostgreSQL 自然語言資料分析 Agent

    A self-hostable natural-language data analysis agent for PostgreSQL

  2. 02/06

    LedgerLion

    記帳與台股提醒Ledger & TW stock reminders

    本地優先的個人記帳與台股公開資訊提醒 Agent

    A local-first personal bookkeeping and Taiwan-stock public-info reminder agent

  3. 03/06

    LensLion

    PR 視覺與無障礙審查PR visual & a11y review

    前端視覺 QA 與無障礙(WCAG)Pull Request 自動審查 Agent

    A frontend visual QA and accessibility (WCAG) pull-request review agent

  4. 04/06

    PulseLion

    Threads 成長分析Threads growth analysis

    Threads 行銷成長與個人演算法分析 Agent

    A Threads marketing-growth and personal-algorithm analysis agent

  5. 05/06

    ProbeLion

    MCP 除錯與評估MCP debugging & eval

    MCP 視覺化除錯工作台與評估沙盒

    A visual debugging workbench and eval sandbox for MCP

  6. 06/06

    RuleLion

    AI 開發規則AI coding rules

    多工具 AI 開發規則產生、事實校驗與同步 CLI

    Generate, verify and sync AI coding rules across tools

(03) 它們怎麼搭在一起How they fit together

一半幫你打造,一半幫你過日子。Half help you build. Half help you live.

這是依各專案 README 的用途分的兩組,不是一條自動串接的流水線——每個 Agent 都能單獨使用。

Two groups by what each README says it is for — not an automated pipeline. Every agent stands on its own.

給開發者For builders

規則 → 工具 → 審查,三者都附 CI 用法。

Rules → tools → review, each with a CI mode.

  1. RuleLionAI 開發規則AI coding rules
  2. ProbeLionMCP 除錯與評估MCP debugging & eval
  3. LensLionPR 視覺與無障礙審查PR visual & a11y review

給日常For everyday work

資料、帳本、社群——資料留在你的環境。

Data, money, audience — data stays with you.

  • QueryLion資料提問Data questions
  • LedgerLion記帳與台股提醒Ledger & TW stock reminders
  • PulseLionThreads 成長分析Threads growth analysis

共通點:BYOK 與遮罩金鑰、系統/淺色/深色主題與鍵盤操作、同一套 Lion family 設計 token。

Shared: BYOK with masked keys, system / light / dark themes with keyboard support, and one set of Lion family design tokens.

(04) 實測數字Measured, not promised

附日期、附測試集、附缺點。Dated. Sourced. Caveated.

全部取自各 repo README,2026-11-02 在各專案自有的示範或測試集上執行,不是第三方評測。

All quoted from the repo READMEs, run on 2026-11-02 against each project’s own demo or test set — not third-party benchmarks.

  • 96.7%

    QueryLion 正確率QueryLion accuracy

    圖表命中Chart hit
    56.7%
    P50 首個 SSE(目標 < 1 s)P50 first SSE (target < 1 s)
    1.6 s

    30 題自有題庫,claude-haiku-5-5,2026-11-02。圖表命中僅 56.7%,首個 SSE 1.6 s 未達 < 1 s 目標。

    Own 30-question set, claude-haiku-5-5, 2026-11-02. Chart hit only 56.7%; first SSE 1.6 s misses the < 1 s target.

    QueryLion/README.md · 2026-11-02

  • 100%

    LensLion 召回率LensLion recall

    自有靶場 12 種已知問題,乾淨頁 0 誤報,加 Vision 每次 PR US$0.0023,2026-11-02。

    12 known issues on its own testbed, 0 false positives on the clean page, US$0.0023 per PR with vision, 2026-11-02.

    LensLion/README.md · 2026-11-02

  • 80%

    ProbeLion 範例題組通過率ProbeLion sample-suite pass rate

    claude-haiku-5-5 80%、claude-sonnet-5-5 75%,自有 20 題,2026-11-02;失敗多來自範例工具只收英文城市名。

    claude-haiku-5-5 80%, claude-sonnet-5-5 75%, own 20 cases, 2026-11-02; most failures trace to the sample tool accepting English city names only.

    ProbeLion/README.md · 2026-11-02

  • 1

    RuleLion 移除的未驗證宣稱Unverified claim removed by RuleLion

    在自己的 repo 上以 claude-haiku-5-5 執行 generate,約 9–10 秒,2026-11-02;單次執行。

    generate with claude-haiku-5-5 on its own repo, ≈ 9–10 s, 2026-11-02; a single run.

    RuleLion/README.md · 2026-11-02

想先試試看?Want an early look?

所有 repo 目前都是私有的,準備開源中。想搶先試用,請透過 GitHub 個人頁聯絡。All repos are private for now and on their way to open source. To try one early, get in touch via the GitHub profile.