cross-account-cost-attribution

name

cross-account-cost-attribution

description

Retrieve, compare, and attribute AWS spend across the two China region accounts (aws-cn in cn-northwest-1 and aws-cn-2 in cn-north-1). Use this skill when the user asks about cost, spend, billing, 花费, 成本, 账单, 多少钱, expensive, 贵, top services, cost breakdown, month-over-month, budget, or wants to know which China account spends more and on what. Covers month-to-date, last month, last 90 days, and custom time ranges. Also use when the user wants to correlate cost with specific resources or services (e.g. "哪个账号的 EC2 花费高"). Report totals per account, top services per account, and deltas vs previous period when available.

Cross-Account Cost Attribution

Routing behavior is governed by china-region-multi-account-routing. This skill assumes the agent is already prepared to hit both China MCPs and focuses on how to query cost data and present a comparative view.

When to use

Use this skill whenever the user's intent involves monetary comparison or attribution across the two China accounts. Examples:

• "这个月两个账号各花了多少"
• "哪个账号贵"
• "两个账号的 EC2 花费对比"
• "北京账号上个月比这个月涨了多少"
• "top 5 services by cost in each account"
• "我这两个中国区账号一共 MTD 多少"

Do not use this skill for:

• Single-account cost queries (the user already specifies account → let the base aws-api-mcp-server handle it)
• AWS global cost (this skill is China-partition only)
• Forecasting / budget alerts (out of scope — different API surface)

China partition cost API note

AWS China has its own Billing/Cost Explorer API surface at bcm-data-exports.cn-northwest-1.amazonaws.com.cn and ce.cn-north-1.amazonaws.com.cn. The CLI commands are the same as global (aws ce get-cost-and-usage), but credentials must be the China-partition AK/SK — which is exactly what each MCP pod already has. Do not use aws-global credentials; they will fail with AuthFailure.

Procedure

Step 1 — Determine the time range

Parse from the user's query. Default if unspecified: month-to-date (MTD) for the current month. Common mappings:

User phrase	Time range
"这个月" / "this month" / "MTD"	First day of current month → today
"上个月" / "last month"	First day → last day of previous month
"过去 30 天" / "last 30 days"	Today − 30 days → today
"今年" / "YTD"	Jan 1 of current year → today
"环比" / "month-over-month"	Return both current MTD and previous full month

Always include the exact date range in the response so the user can verify.

Step 2 — Query Cost Explorer on both endpoints in parallel

Use aws ce get-cost-and-usage with:

• --time-period Start=...,End=...
• --granularity MONTHLY (or DAILY if range ≤ 31 days and user wants trend)
• --metrics UnblendedCost
• --group-by Type=DIMENSION,Key=SERVICE (for top-services breakdown)

Call on both aws-cn and aws-cn-2 MCPs concurrently.

Step 3 — Aggregate and compare

For each account, compute:

1. Total spend (sum of UnblendedCost.Amount across groups)
2. Top N services (default N=5, sorted by amount desc)
3. Delta vs prior period (if the user asked for month-over-month or if the default 1-month view makes comparison obvious)

Step 4 — Present

Two-part format. Headline numbers first, then breakdown tables.

Headline:

Date range: 2026-05-01 → 2026-05-11 (MTD, 11 days)

aws-cn   (Ningxia): ¥ 1,234.56  (ΔMoM: +8% vs April)
aws-cn-2 (Beijing): ¥   456.78  (ΔMoM: −3% vs April)
Combined:           ¥ 1,691.34

Top-services table (side-by-side):

Service	aws-cn	aws-cn-2
Amazon EC2	¥ 800	¥ 200
Amazon RDS	¥ 300	¥ 150
Amazon S3	¥ 80	¥ 50
Data Transfer	¥ 30	¥ 30
Other	¥ 24	¥ 26
Total	¥ 1,234.56	¥ 456.78

Step 5 — Interpret (1–3 sentences)

State one observation: where is the bigger account spending, is the delta normal, is there an outlier service. Do not recommend cost optimization unless asked — that requires deeper data than this skill retrieves.

Things not to do

• Do not convert currencies. China region returns CNY; present as ¥ and do not add a $ equivalent. Exchange rate assumptions cause confusion and you do not have a rate source.
• Do not compare different time ranges across the two accounts. If one account's data is incomplete (e.g. CE has 24h lag), note the discrepancy explicitly instead of silently aligning.
• Do not use --metrics BlendedCost or AmortizedCost without the user asking. Default to UnblendedCost — it matches what users see on the billing console.
• Do not claim a spending anomaly unless you have ≥2 prior periods for comparison. One-month deltas on new accounts are noisy.
• Do not include detailed SKU-level breakdown unless the user asks. Service-level is the default granularity.

Examples

Input: "这两个账号这个月花了多少"

Action: MTD on both in parallel, service-level breakdown, side-by-side table. Report total per account + combined.

Input: "哪个账号贵，贵在哪"

Action: Current MTD on both, find the higher one, break down its top-3 services, highlight which service drives the difference.

Input: "北京账号 EC2 比上月涨了没"

Action: aws-cn-2 only, current MTD vs previous full month, filter service=Amazon EC2. Report delta in absolute ¥ and percent. Note if the current month is partial (MTD) and adjust interpretation accordingly.

Input: "过去 90 天两个账号哪天花的最多"

Action: Both accounts, daily granularity, past 90 days. Find the peak-spend day per account and report with date + amount + top service that day.