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

Wiki Management

Usage Examples

Once the server is connected, talk to Claude Code naturally:

Ingest a URL:

“Read this article and add it to the wiki: https://example.com/interesting-article”

Ingest a PDF:

“Extract this paper and create wiki pages from it: /path/to/paper.pdf”

Transcribe a lecture:

“Transcribe this lecture and summarize the key points into wiki pages: /path/to/lecture.mp4”

Search your knowledge base:

“What do I have in my wiki about transformer architectures?”

Maintain the wiki:

“List all pages and check for any that are orphaned or could use cross-links”

Wiki Conventions

• Page names: kebab-case.md (e.g., transformer-architecture.md)
• Frontmatter: YAML between --- delimiters with at minimum title, created, updated, tags
• Links: Obsidian-style [[page-name]] wikilinks
• Index: wiki/index.md — master table of contents organized by category
• Log: wiki/log.md — append-only changelog with timestamps

Example page:

---
title: Transformer Architecture
created: 2026-04-09
updated: 2026-04-09
tags: [deep-learning, nlp, architecture]
source: https://arxiv.org/abs/1706.03762
---

# Transformer Architecture

The transformer is a neural network architecture based entirely on
[[attention-mechanism|attention mechanisms]], dispensing with recurrence
and convolutions entirely.

## See Also

- [[self-attention]]
- [[bert]]
- [[gpt]]

Configuration

All configuration is via environment variables or application.yml overrides:

Project Structure

src/main/kotlin/com/wiki/agent/
├── WikiAgentApplication.kt       # Spring Boot entry point
├── config/WikiProperties.kt      # Configuration properties
├── tool/
│   ├── ExtractionTools.kt        # wiki_extract, wiki_transcribe
│   └── WikiTools.kt              # Wiki CRUD + search tools
├── service/
│   ├── WikiService.kt            # Filesystem operations
│   └── ExtractionService.kt      # Extraction dispatch
├── extract/
│   ├── Extractor.kt              # Interface
│   ├── UrlExtractor.kt           # Jsoup HTML → text
│   ├── PdfExtractor.kt           # Tika PDF → text
│   ├── TextExtractor.kt          # Passthrough .md/.txt
│   ├── AudioExtractor.kt         # Whisper API
│   └── VideoExtractor.kt         # ffmpeg → Whisper
└── model/SourceDocument.kt       # Extraction result data class

Development

./gradlew test          # Run tests (11 tests)
./gradlew bootJar       # Build fat JAR
./gradlew compileKotlin # Compile only

Note: The project requires JDK 21 for both building and running. If your default java points to a different version, set JAVA_HOME or update gradle.properties:

org.gradle.java.home=/path/to/jdk-21

Architecture Notes

• The MCP server makes zero LLM calls — all intelligence lives in Claude Code
• Communication is STDIO only (spring.main.web-application-type=none)
• All logging goes to logs/wiki-agent.log, never stdout (stdout is reserved for MCP protocol)
• Raw sources in raw/ are immutable — they preserve the original extraction with metadata
• The wiki in wiki/ is a standard Obsidian vault and can be opened directly

License

MIT