Claude Code 와 함께 ECI CLI 사용하기
Claude Code 에서 ECI CLI를 자연어로 다루기 위한 SKILL을 제공합니다.
Claude Code 전용
Cursor · Windsurf 등 다른 AI 도구를 쓰신다면 본문 내용을 해당 도구의 컨벤션에 맞춰 복붙해서 사용해 주세요.
설치
아래 코드블록 내용을 그대로 복사해서 ~/.claude/skills/eci-cli/SKILL.md 로 저장하세요.
---
name: eci-cli
description: Use the `eci` CLI to manage Elice Cloud Infrastructure (ECI) resources. Invoke when the user mentions ECI, the `eci` command, or asks to create/manage VMs, networks, or storage on the ECI platform — including GPU/NPU clusters (e.g., H100, B200), block/object storage, and public IPs.
---
# ECI CLI
`eci` manages compute, network, and storage on ECI (Elice Cloud Infrastructure).
## Source of truth
This skill is a snapshot of conventions and rules. **Always run `<command> --help` first** for destructive or unfamiliar commands (`launch`, `delete`, `attach`, `resize`). If this skill and `--help` disagree, `--help` wins. For values that change over time (prices, instance types, image versions, status names), query the CLI live rather than relying on this skill.
## Preflight
```bash
eci config verify
```
If it fails, guide the user through `eci config set ...` (preferred over interactive `init` in agent contexts). Don't touch resources before this passes.
## Patterns
- Resource commands follow `eci <group> <resource> list | get <name|id>`
- Names and UUIDs are interchangeable
- `--format {table,json,csv}` + `--query col1,col2` for parsing
- JSON output auto-resolves `*_id` fields to names
## VM launch
```bash
eci compute vm launch \
--name <name> --instance-type <type> --password <pw> [--wait]
```
Most fields have sensible defaults when omitted (image and size depend on instance family; subnet auto-creates a default vnet/subnet on first use). Run `eci compute vm launch --help` for the current flags and default values.
Common variants: `--no-start` (skip boot — $0 compute), `--no-public-ip`, `--no-network`, `--price-type spot` (accelerator types only), `--spec <name>`, `--pricing-id <uuid>` (when multiple pricings exist for one instance type).
## Safety rules — must follow
1. **`vm delete` always with `--cascade`** — without it, the disk / NIC / Public IP leak and keep billing.
2. **Bulk operations require explicit confirmation.** List targets first.
3. **Never delete vnet / subnet / disk in use** without checking attachments.
4. **Password rules** (API-enforced — confirm exact rules via the launch error message or `--help`): typically minimum length, multiple character classes, and no long ascending/descending sequence.
5. **CIDR ranges**: only specific private ranges are allowed (10.x is rejected). Confirm the exact allowed ranges from the API error if a `vnet create` fails. Subnet `--gateway` requires CIDR notation, not a bare IP.
## Cost awareness
State the hourly cost **before** showing the launch command, especially for GPU/NPU instances. **Get current prices from the CLI rather than assuming:**
```bash
eci pricing list --resource-kind vm_allocation
eci pricing list --resource-kind public_ip
eci pricing list --resource-kind block_storage
```
For batches (3+ VMs or any GPU/NPU VM), show total cost + cleanup plan up front. `idle` VMs cost 0 for compute (disk + IP still apply) — don't auto-`start` unless asked.
For any ondemand/reserved VM launch, verify quota first:
```bash
eci org info --format json
```
Check `resource_quota.compute.devices`, `instance_types.<uuid>`, `storage.*`, `network.public_ip`. State the gap if over and offer to reduce count or request a quota increase. Pre-checking lets you explain *why* a launch will fail before the API rejects it, which is clearer than reacting to the error after the fact.
**Spot pricing bypasses quota.** `--price-type spot` launches do not consume `resource_quota.compute.devices` or `instance_types.<uuid>` — do not block a spot launch (single or batch) on these fields being absent or zero. Quota only constrains ondemand/reserved. Storage and `network.public_ip` quota still apply regardless of pricing type.
## Known behaviors
- VM/disk status names differ from AWS/GCP (`running`/`stopped`). Read the current state via `vm get`/`block get` rather than assuming.
- `vm stop` → idle transition can take 1–3 minutes. A follow-up `vm start` only succeeds once the VM has fully reached the idle state.
- After a disk resize, the disk briefly returns to a transient state. `vm start` is rejected until the disk settles back to its ready state.
- `vm delete --cascade` can timeout at the stop phase. Retry; for bulk cleanup use `|| true` in a loop.
## Don'ts
- Don't run `eci` silently — confirm what gets created/destroyed first.
- Don't drop `--cascade` as "safer" — it's the opposite.
- Don't guess a single `--pricing-id` when multiple exist; surface the API's candidate list to the user.
- Don't add `--wait` to GPU launches unless asked (each adds 1–3 min of blocking).
- Don't quote specific prices, quota numbers, or instance type lists from memory — fetch live via `eci pricing list`, `eci instance-type list`, and `eci org info`.
동작 확인
설치 후 Claude Code를 새 세션으로 열고 다음과 같이 시도해 보세요:
ECI에 C-2 가상머신 1대 띄워줘
Claude가 다음 순서로 응답하면 skill이 올바르게 작동하는 것입니다:
- 시간당 비용을 먼저 명시 (
eci pricing list로 확인한 값) eci config verify·eci instance-type list등으로 환경 확인- 사용자 확인을 받은 뒤
eci compute vm launch ...제안 - 작업 후
vm delete --cascade정리 안내
skill이 포함하는 내용 요약
| 영역 | 내용 |
|---|---|
| 명령 패턴 | list/get 통일, launch 자주 쓰는 플래그, SSH, lifecycle, delete |
| 안전 규칙 | --cascade 강제, bulk 확인, 비밀번호 규칙, CIDR 규칙 |
| 비용 인식 | 자원별 단가표, GPU launch 전 비용 명시 |
| 알려진 동작 | stop → idle 1~3분, resize 후 disk queued, cascade timeout 재시도 패턴 |
| 하지 말 것 | silent 실행, --help 없이 명령 조립, pricing-id 추측 |
참고
- ECI CLI 매뉴얼: github.com/elice-dev/eci-cli
- Claude Code skill 시스템: code.claude.com/docs/ko/skills