Agent Skills › langfuse/langfuse › analyze-cloud-costs

analyze-cloud-costs

GitHub

基于 Metabase 成本数据源，分析 Langfuse Cloud 的支出结构、驱动因素及异常回归。通过查询成本集市提供按时间、提供商和服务维度的证据支持报告。

.agents/skills/analyze-cloud-costs/SKILL.md langfuse/langfuse

Trigger Scenarios

询问云支出或成本结构分析 AWS 与 ClickHouse 成本拆分排查成本回归或异常增长

Install

npx skills add langfuse/langfuse --skill analyze-cloud-costs -g -y

More Options

Non-standard path

npx skills add https://github.com/langfuse/langfuse/tree/main/.agents/skills/analyze-cloud-costs -g -y

Use without installing

npx skills use langfuse/langfuse@analyze-cloud-costs

指定 Agent (Claude Code)

npx skills add langfuse/langfuse --skill analyze-cloud-costs -a claude-code -g -y

安装 repo 全部 skill

npx skills add langfuse/langfuse --all -g -y

预览 repo 内 skill

npx skills add langfuse/langfuse --list

SKILL.md

Frontmatter

{
    "name": "analyze-cloud-costs",
    "description": "Analyze Langfuse Cloud infrastructure cost structure using Metabase cost\nmarts. Use when asked about cloud spend, AWS versus ClickHouse cost splits,\ncost drivers by provider\/service\/usage type\/account, daily cost per tracing\nevent, infra cost dashboards, or cost regressions visible in Metabase."
}

Analyze Cloud Costs

Overview

Use this skill for evidence-backed Langfuse Cloud cost analysis. The primary source is the Metabase infra cost dashboard and its production cost marts; the deliverable should name the time window, query grain, top drivers, and caveats.

Workflow

Clarify the question and choose the grain:
- Headline daily totals: total, AWS, ClickHouse, tracing events, and cost per 100k events.
- Cost structure: provider, service, usage type, operation, account, and day.
- Driver or regression analysis: compare a recent complete-day window against a prior baseline.
Load references/cost-marts.md for table IDs, field IDs, query examples, and caveats.
Use the Metabase MCP. If the Metabase tools are not visible, discover them with tool search before falling back to manual interpretation.
Prefer complete UTC days. Avoid treating current-day AWS cost as final because AWS CUR rows can arrive late.
Start broad, then drill down:
- Provider split.
- Service split within the dominant provider.
- Usage type, operation, and account split for the top services.
- Daily trend when explaining change over time.
Report only what the queried data supports. If a requested slice is absent, say that no rows were found for that slice instead of inventing a driver.

Query Rules

Use mcp__metabase__.query for quick reads. Use construct_query plus execute_query when you need to inspect or reuse the opaque query.
Pass filters, aggregations, group_by, and fields as JSON arrays. Some tool schemas may display these as strings; if that happens, serialize the same arrays without changing their shape.
Keep limits explicit and small enough for analysis. Use pagination only when the continuation token is needed.
Include the Metabase dashboard link or query result context in the final answer when useful.

Output Expectations

Summarize:

Time window and whether it uses complete UTC days.
Total cost and provider split when relevant.
Top cost drivers by service, usage type, operation, or account.
Trend or baseline comparison when the user asks "why did this change?"
Caveats, especially incomplete current-day AWS data and ClickHouse credit labeling in the unified mart.