这是一个基于 markdown 的技能，教 Claude Code 如何通过 REST API 与 Elasticsearch 集互。没有 SDK，没有客户端库，没有 MCP 服务器——只有 curl 和知识。

里面有什么

安装

Claude Code 是从本地目录加载技能，而不是从 GitHub URL 加载。克隆这个仓库，然后将文件复制到支持的技能位置之一。

选项A：个人技能（所有项目均可用）

选项B：项目技能（仅在一个仓库中提供）

在你项目的根目录中：

然后把.claude/skills/elasticsearch/目录提交到你的仓库，这样你的团队也能拿到。

验证

重启Claude代码，然后检查技能是否已加载：

你应该会在列表中看到elasticsearch。

Elasticsearch Skill for Claude Code

A markdown-based skill that teaches Claude Code how to interact with Elasticsearch clusters via REST API. No SDK, no client library, no MCP server — just curl and knowledge.

What’s in it

Installation

Claude Code loads skills from local directories, not from GitHub URLs. Clone this repo, then copy the files to one of the supported skill locations.

Option A: Personal skill (available in all projects)

git clone https://github.com/davidgeorgehope/elasticsearch-skill.git
mkdir -p ~/.claude/skills/elasticsearch
cp elasticsearch-skill/SKILL.md ~/.claude/skills/elasticsearch/
cp -r elasticsearch-skill/references ~/.claude/skills/elasticsearch/

Option B: Project skill (available in one repo only)

From within your project’s root directory:

git clone https://github.com/davidgeorgehope/elasticsearch-skill.git /tmp/elasticsearch-skill
mkdir -p .claude/skills/elasticsearch
cp /tmp/elasticsearch-skill/SKILL.md .claude/skills/elasticsearch/
cp -r /tmp/elasticsearch-skill/references .claude/skills/elasticsearch/

Then commit the .claude/skills/elasticsearch/ directory to your repo so your team gets it too.

Verify

Restart Claude Code, then check the skill is loaded:

/skills

You should see elasticsearch in the list.

Note: Adding a GitHub URL to the skills array in ~/.claude/settings.json does not work. Skills must be local files — there is no remote fetching.

Configuration

Set your cluster credentials as environment variables before starting Claude Code:

export ES_URL="https://your-cluster.es.cloud.elastic.co:443"
export ES_API_KEY="your-base64-api-key"

Optionally, for Kibana API access:

export KIBANA_URL="https://your-cluster.kb.cloud.elastic.co:443"

No server to run, no dependencies to install, no Docker container to spin up.

Why a Skill and Not MCP?

This is a deliberate choice, not a limitation.

The token tax

MCP servers expose tools via JSON schemas that get injected into the LLM’s context window on every request. Anthropic’s own testing found that a modest five-server setup with 58 tools consumed ~55,000 tokens before the conversation even started. That’s your context window being eaten alive by tool definitions you might not even use.

A skill loads its main instructions only when triggered, and deeper reference files only when the LLM actually needs them. Progressive disclosure by design — not by workaround.

The full circle