
MCP服务器将原始源代码转化为一个用Claude代码连接的Obsidian维基。
概览
一个个人知识管理系统,将原始资源——URL、PDF、音频、视频——转化为互联维基。它作为 Kotlin + Spring Boot 的 MCP 服务器构建,并与 Claude Code 配合:服务器是工具包,Claude 是大脑。 灵感来自Andrej Karpathy利用大型语言模型逐步构建个人维基,而非RAG的模式。 你给 Claude Code 一个 URL、PDF 或音频文件。它调用服务器提取内容,然后用自身智能将其提炼成结构完善、交叉链接和元数据的维基页面。维基是一个兼容 Obsidian 的保险库,你可以浏览和搜索。 - — Amazon Corretto 或任何 JDK 21+ - — 安装指南 - (可选)— 仅用于视频转录(brew install ffmpeg)- (可选)— 仅音频/视频转录需要 这会产生构建/库/wiki-agent.jar(~70MB,包含Apache Tika)。 使用 -s 用户让服务器在所有项目中可用。仅在当前项目中使用 -s 项目。
概述(AI 翻译)
一个个人知识管理系统,将原始资源——URL、PDF、音频、视频——转化为互联维基。它作为 Kotlin + Spring Boot 的 MCP 服务器构建,并与 Claude Code 配合:服务器是工具包,Claude 是大脑。
灵感来自Andrej Karpathy利用大型语言模型逐步构建个人维基,而非RAG的模式。
工作原理
你给 Claude Code 一个 URL、PDF 或音频文件。它调用服务器提取内容,然后用自身智能将其提炼成结构完善、交叉链接和元数据的维基页面。维基是一个兼容 Obsidian 的保险库,你可以浏览和搜索。
先修条件
- Java 21+ — Amazon Corretto 或任何 JDK 21+
- Claude Code — 安装指南
- ffmpeg(可选)——仅在视频转录时要求(
brew install ffmpeg) - OpenAI API 密钥(可选)——仅用于音频/视频转录
快速启动
1.建造
这会产生 build/libs/wiki-agent.jar(~70MB,包含Apache Tika)。
2.用Claude代码注册
使用 -s user 使服务器在所有项目中可用。仅使用 -s project 用于当前项目。
3.配置路径(推荐)
本页由 AI 翻译,技术细节请以原文为准。
README
以下为原文 README 节选,完整内容请查看原文。
LLM Wiki Agent
A personal knowledge management system that turns raw sources — URLs, PDFs, audio, video — into an interlinked wiki. Built as an MCP server in Kotlin + Spring Boot, it pairs with Claude Code: the server is the toolkit, Claude is the brain.
Inspired by Andrej Karpathy’s pattern of using LLMs to incrementally build a personal wiki instead of RAG.
How It Works
Claude Code (LLM — all intelligence)
│ MCP protocol (STDIO)
▼
Wiki Agent Server (toolkit — zero LLM calls)
├── Content extraction (URL, PDF, text, audio, video)
├── Wiki management (CRUD, search, index, log)
│
▼ filesystem
wiki/ (Obsidian vault) raw/ (immutable sources)
You give Claude Code a URL, PDF, or audio file. It calls the server to extract content, then uses its own intelligence to distill it into wiki pages with proper structure, cross-links, and metadata. The wiki is an Obsidian-compatible vault you can browse and search.
Prerequisites
- Java 21+ — Amazon Corretto or any JDK 21+
- Claude Code — Installation guide
- ffmpeg (optional) — required only for video transcription (
brew install ffmpeg) - OpenAI API key (optional) — required only for audio/video transcription
Quick Start
1. Build
git clone && cd llm_wiki_agent
./gradlew bootJar
This produces build/libs/wiki-agent.jar (~70MB, includes Apache Tika).
2. Register with Claude Code
claude mcp add -s user wiki-agent \
-- java -jar /absolute/path/to/llm_wiki_agent/build/libs/wiki-agent.jar
Use -s user to make the server available in all projects. Use -s project for the current project only.
3. Configure paths (recommended)
By default, the wiki and raw directories are created relative to the working directory. For a persistent setup, set absolute paths in ~/.claude.json under the mcpServers entry:
{
"mcpServers": {
"wiki-agent": {
"type": "stdio",
"command": "java",
"args": ["-jar", "/absolute/path/to/wiki-agent.jar"],
"env": {
"WIKI_PATH": "/absolute/path/to/wiki",
"RAW_PATH": "/absolute/path/to/raw",
"OPENAI_API_KEY": "sk-..."
}
}
}
}
4. Verify
claude mcp list
# wiki-agent: java -jar ... - ✓ Connected
5. Open in Obsidian (optional)
Open your wiki/ directory as an Obsidian vault to get a graph view of all your pages and their [[wikilinks]].
MCP Tools
Content Extraction
| Tool | Description |
|---|---|
wiki_extract |
Extract content from a URL, PDF, or text file. Auto-detects source type. Saves raw extraction to raw/. |
wiki_transcribe |
Transcribe audio/video via OpenAI Whisper. For video, extracts audio with ffmpeg first. Requires OPENAI_API_KEY. |
Wiki Management
安装
This server does not publish a one-line install command.
Open the repository installation guide 新手?先看 mcpskill.org 教程 →配置
{
"mcpServers": {
"wiki-agent": {
"type": "stdio",
"command": "java",
"args": ["-jar", "/absolute/path/to/wiki-agent.jar"],
"env": {
"WIKI_PATH": "/absolute/path/to/wiki",
"RAW_PATH": "/absolute/path/to/raw",
"OPENAI_API_KEY": "sk-..."
}
}
}
}