CRM Lifecycle Ops

io.github.har5h98由 GitHub 帳號 har5h98 發布v0.5.4更新於 Nov 7, 2026

AI copilot for lifecycle and CRM ops on Braze, Klaviyo and Iterable. Read-only.

已驗證STDIO僅桌面效率與工作流商業與電商

概覽

AI 產生的概覽

針對 Braze、Klaviyo 與 Iterable 的唯讀生命週期與 CRM 營運助手,可規劃活動建置並在寄送前檢查訊息。

可以做什麼
  • 規劃與建置活動:依平台術語提供逐步建置說明、含重疊檢查的旅程設計、批次尋找取代預覽
  • 評估變更影響:屬性、分群、區塊、事件或連結變更會影響什麼,自上次以來的變化,安全修改方案,旅程圖示
  • 依管道檢查訊息:電子郵件、推播、簡訊、應用程式內訊息、內容卡片與 webhook,並依邊界情況客戶渲染個人化內容
  • 檢查整個活動與旅程:受眾、轉換事件、命名規範、斷頭分支、無法到達的步驟
  • 營運報告:上線後送達健康度對照門檻、檢查次數與發現問題數,以及把問題回報轉為測試案例
適用情境
適合手動在 Braze、Klaviyo 或 Iterable 中建置活動的生命週期、CRM 與行銷營運團隊,用於寄送前複核、變更影響分析或建置指引。內建示範帳戶,無需連接任何平台即可試用。
執行需求
透過 npx 在本機執行,需要 Node.js 20+;README 以 Claude Code 為主,但任何 MCP 用戶端都能執行此引擎。連接平台需在 CLO_BRAZE_API_KEY、CLO_KLAVIYO_API_KEY 或 CLO_ITERABLE_API_KEY 提供唯讀 API 金鑰,並設定 CLO_BRAZE_REST_ENDPOINT、CLO_ITERABLE_REGION 與帳戶標籤等。可選 CLO_RULES 與 CLO_TEAM_DIR。不設金鑰即使用示範帳戶。
安裝前請注意
處於早期 alpha:僅在示範帳戶與模擬 API 回應上驗證,尚未在真實帳戶實測。它會以唯讀金鑰讀取真實帳戶資料,並在要求時對外部連結發出檢查。它會在使用者家目錄下寫入快照、團隊規則與使用日誌;若設定 CLO_TEAM_DIR,還會複製僅含計數的使用日誌。經確認後 /report-problem 可提交去識別化的 GitHub issue。它不會在你的帳戶中寄送或修改任何內容。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 CRM Lifecycle Ops,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

crm-lifecycle-ops

Pre-send QA for Braze, Klaviyo and Iterable. Every email, text, push and in-app message is checked by code, as every kind of customer, before it goes out. Then, in your live account: what a change will break, and how to build it.

[CI] [Release] [Status: early alpha] [License: MIT]

[Terminal showing /check-campaign flagging a staging link and a blank first name, /what-breaks listing live programs that use loyalty_tier, and /edit-plan warning that a footer block is shared by three other live programs before giving Braze steps]

A Claude Code plugin (and a CLI for CI) for lifecycle, CRM and marketing-ops teams who build campaigns by hand. The message checks work fully today; the live-account features are early alpha (see the status column below). It is read-only: it never sends or edits anything.

Try it in 30 seconds

No account and no Claude Code needed, just Node.js 20+. This checks four sample emails that ship with the package:

bash
npx crm-lifecycle-ops check --demo

[Terminal running npx crm-lifecycle-ops check --demo: one sample email is ready, one links to a staging site and must not be sent, one greets people without a name as 'Hi ,', and one has accessibility problems and a 'View in browser' inbox preview]

Then point it at your own templates: npx crm-lifecycle-ops check path/to/emails. That covers emails. The plugin below adds the rest: every channel, your live account, what a change will break, and build steps.

What it does

Try
Plan and build fasterClick-by-click build steps in your platform's terms, with your naming and UTMs/build-plan
Change safelySee what a change to a segment, field or shared block will break, before you make it/what-breaks
Send with confidenceEvery message on every channel, checked as every kind of customer/check-campaign
Prove the impactDelivery health after launch, plus problems caught and hours saved/ops-report

Quick start

Needs Claude Code and Node.js 20+.

/plugin marketplace add har5h98/crm-lifecycle-ops/plugin install crm-lifecycle-ops@crm-lifecycle-ops

Then try it on the built-in demo account. No platform access needed:

  1. /refresh demo loads a fictional brand's Braze workspace.
  2. /check-campaign Price Drop Alert checks a whole campaign.
  3. /what-breaks loyalty_tier shows what a change would affect.

When you're ready for your own account:

  • /setup teaches it your team's rules (about five minutes, or paste an email you like and it learns from that).
  • /connect adds a read-only key. Keys go in the plugin's settings and your system's secure storage, never in chat.
Using Cursor, Claude Desktop or another MCP client?

The engine is also on npm, so any MCP client can run it. The slash commands are Claude Code skills, so elsewhere you ask in plain words ("check this email", "what breaks if I rename loyalty_tier?").

json
{  "mcpServers": {    "crm-lifecycle-ops": {      "command": "npx",      "args": ["-y", "crm-lifecycle-ops"],      "env": { "CLO_BRAZE_API_KEY": "…", "CLO_BRAZE_REST_ENDPOINT": "https://rest.iad-01.braze.com" }    }  }}

Leave env out to use the demo account. Every setting is listed in server.json.

Check email templates in CI

The same checks run from a terminal, so a pipeline can stop a broken email before anyone launches it. It exits 1 on a blocking problem (or on warnings too, with --fail-on warn). It uses the built-in rules unless you pass a rules file, so it gives the same verdict on every machine; commit your team's file (the one /setup saves to ~/.crm-lifecycle-ops/rules.json) and point at it:

bash
npx crm-lifecycle-ops check emails/ --rules team-rules.json
yaml
# .github/workflows/email-qa.ymlon: pull_requestjobs:  email-qa:    runs-on: ubuntu-latest    steps:      - uses: actions/checkout@v4      - run: npx -y [email protected] check emails/ # pin a version so new checks don't change your build

--help lists the options (--platform, --subject, --transactional, --json). The template language is detected per file; the output says "guessed" when nothing in a file gives it away, so pass --platform then.

Early alpha. Everything works on the demo account and against mocked API responses. It hasn't been field-tested on real accounts yet. If a result looks wrong, /report-problem turns it into a test and a fix.

Example

Real output from the demo account:

/check-campaign Price Drop Alert          🔴 Don't send. 1 blocking problem
fail  links.non-production-host        "See items" points at staging.fernbrook.examplewarn  personalization.dangling-punct.  Someone with no first name on file sees "Hi ,"warn  html.img-missing-alt             1 image has no alt textwarn  a11y.no-lang                     No lang attribute, so screen readers guess the language
/check-sms Flash Sale Text                🟡 Fix before sending
warn  sms.unicode            "’" and "—" switch the text to Unicode: 70 characters per segment, not 160.                             Swap ’ → ' and — → -warn  sms.too-many-segments  Up to 3 billed segments; house limit is 2warn  sms.no-opt-out         No "Reply STOP to opt out"

Code makes every pass/fail call, never AI. Claude adds a labelled, advisory review (subject line, CTA, tone) on top. Need a record for the launch ticket? Ask /check-campaign for a QA sign-off report: a Markdown file with every finding, what couldn't be verified and a reviewer sign-off, ready to paste into Jira, Slack or Notion.

All commands

✅ works fully today · 🧪 works on the demo account and mocked APIs, not yet field-tested on real accounts

CommandWhat it doesStatus
Get started/start-hereOverview and the best first step✅
/setupTeach it your team's rules: language, brand, links, footer, senders, test customers, strictness✅
/house-rulesView or change those rules; mark a finding "fine for us"✅
/connectConnect Braze, Klaviyo or Iterable with a read-only key🧪
/refreshPull the latest campaigns, journeys, segments, templates and content blocks (or load the demo)🧪
Plan and build/plan-journeyDesign a journey and check it for overlap with what's already live🧪
/build-planStep-by-step build instructions in your platform's own terms🧪
/bulk-editPreview a find-and-replace across every subject, body and content block🧪
Change safely/what-breaksWhat's affected if you change or delete an attribute, segment, block, event or link🧪
/what-changedWhat changed in the account since last time, live items first🧪
/edit-planThe safe way to change something live (copy, running journeys, shared blocks, audiences)🧪
/show-journeyDraw a live journey as a diagram, with structural problems flagged🧪
Send with confidence/check-campaignA whole campaign or journey: every message, settings and structure🧪
/check-emailOne email: HTML, links, unsubscribe, UTMs, personalization, house rules✅
/check-pushOne push: lock-screen length for your longest names, deep links, blanks✅
/check-smsOne text: billed segments, Unicode traps, opt-out wording, sender name✅
/check-in-appOne in-app message or content card: length, buttons, links✅
Prove the impact/watch-launchBounce, unsubscribe, spam and delivery rates against thresholds🧪
/ops-reportMessages checked, problems caught, hours saved, for you or the whole team✅
/report-problemReport a wrong result; it becomes a test case and gets fixed✅

Details

What "check" covers, by channel
ChannelCatches
EmailBroken Outlook blocks, unclosed tables, missing alt text, Gmail clipping; a preheader that's missing, repeats the subject or reads "View in browser"; staging or malformed links, http://, inconsistent UTMs, broken pages (live check); missing unsubscribe (skipped for transactional mail)
Accessibility (email)Missing lang, "click here" links, image-only links with no alt text, layout tables screen readers read as data, text under 12px
PushTitle and body length for your longest customer names, line breaks that hide titles on iOS, deep links, Android's required title (Braze)
SMSBilled segments the way carriers count them, the exact character that switches a text to Unicode, opt-out wording, sender name, each platform's limits
In-appBody length, buttons and their links, click-through destination, HTML structure
Content cardsTitle and message length, card link and image destinations
WebhooksMissing or staging URLs, keys written into the URL, JSON bodies that break when a name contains a quote
Personalization (all)Renders every message for a complete, empty, long-values and accented-name customer (or your own): Hi , blanks, literal null, leftover {{ }}, template errors, Iterable comparisons that silently skip the send
Campaigns & journeysAudience and conversion event, naming convention, dead-end branches, unreachable steps, empty message steps

Braze Liquid, Klaviyo's template language and Iterable Handlebars are supported. Platform tags that can't run locally (coupon codes, catalog lookups, Connected Content) are reported as not verified, never as passing or failing. Full list with ids: src/checks/catalog.ts.

Your rules, not just ours

What's "wrong" depends on the team, so checks come in three layers:

LayerExamplesCan a team change it?
Locked: wrong for everyoneNo unsubscribe, template syntax that doesn't render, broken linksNo
Defaults: sensible everywhereMissing alt text, http:// links, Gmail clipping, mixed UTM taggingRe-grade or turn off
House rules: your conventionsSpelling, date format, brand terms, banned phrases, approved domains and senders, UTM naming, footer, subject length, name fallback, naming convention, test customersSet in /setup

"That's fine for us" becomes a recorded exception with a reason (/house-rules), which is different from "the tool is wrong" (/report-problem). Teams share one rulebook by setting the same Shared rules file in the plugin's settings.

What it can't see (yet)

Every result says what it couldn't verify, and every dependency says how it's known (confirmed by the API, read from the template, read from the segment description, seen on screen).

  • Local rendering is an approximation. Braze Connected Content and catalog lookups aren't executed, and Klaviyo's template language is approximated. These are reported as not verified.
  • Braze's API doesn't say which segment a campaign or Canvas targets. /what-breaks lists those as "can't rule out" and can read them from your Braze dashboard via Claude in Chrome, with your OK.
  • Iterable's API doesn't return journey step order, so Iterable journeys show their emails but not the path.
  • Braze content cards come back from the API with few fields; card checks are fullest when the copy is pasted.
  • It never builds or edits in your account. /build-plan and /edit-plan tell you how; /bulk-edit previews only.

What it runs, sends and fetches

It never sends messages, never edits your account, never reads customer profiles (it tests with made-up customers), and sends nothing to the author or any third party.

WhatWhenWhere it goes
A local MCP server (node dist/mcp-server.js, readable source in src/)When Claude Code loads the pluginYour machine only
Read-only API calls/refresh, /watch-launchOnly the platform you connected: your Braze REST endpoint, a.klaviyo.com, or api.iterable.com / api.eu.iterable.com
Link checks (unsubscribe and tracking links are never requested)Only when you ask for a live link checkThe links in the email being checked
Local files: snapshots, house rules, QA sign-off reports, a usage log of check ids and countsNormal use~/.crm-lifecycle-ops/
A copy of the usage log (counts only, never message content)Only if you set a team report folderThe folder you chose
A GitHub issueOnly via /report-problem, after you approve the exact textgithub.com/har5h98/crm-lifecycle-ops

API keys are read only from the plugin's settings (secure storage). See SECURITY.md and the privacy policy.

Learn more

  • Operator guide: a one-page routine for setup, each step of a build, and the weekly review.
  • How it works: the architecture, data flow and design decisions in plain language.

How it gets better

  1. You report: /report-problem turns a wrong result into a redacted GitHub issue (it asks before posting).
  2. It becomes a test: each confirmed report joins tests/corpus/ and must pass from then on.
  3. Fixes are drafted automatically: Claude writes the test and the fix as a pull request for human review.
  4. Vendor changes are watched: a weekly job checks the Braze, Klaviyo and Iterable APIs for changes, and a monthly one opens a checklist to re-verify the build and edit guides against the vendors' docs. Every plan says when its steps were last verified, and warns once that's more than 90 days ago.

Wrong results are the most valuable contribution. See CONTRIBUTING.md.

License

MIT. Not affiliated with or endorsed by Braze, Klaviyo or Iterable; product names are trademarks of their owners.

來源:README.md,提交 9c2450b

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

2
  1. v0.5.4最新Nov 7, 2026
  2. v0.5.1Nov 3, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分