Overview
llm-wiki-agent (SamurAIGPT) is a 3.3k⭐ agent skill for the LLM Wiki pattern with multi-model support. Drop source documents into raw/ and tell the agent to ingest — it reads them, extracts knowledge, and builds a persistent interlinked wiki. Works with Claude Code, Codex, OpenCode, Gemini CLI.
Key Facts
- Stars: ~3.3k
- Form: agent skill (multi-model — not locked to one LLM)
- Role: packaged agent that operates an LLM Wiki across models
- License: MIT
Features
Core Architecture
- Agent skill form factor — drops into Claude Code, Codex, OpenCode, or Gemini CLI via config files (CLAUDE.md, AGENTS.md, GEMINI.md)
- Plain markdown + git — no server, no database, runs entirely locally; version history for free
- Three-layer output — raw/ (immutable sources), wiki/ (LLM-maintained pages), graph/ (visualization)
Ingest & Processing
- Multi-format ingest — auto-converts PDF, DOCX, PPTX, XLSX, HTML, EPUB, CSV, JSON, XML, and more via markitdown
- Batch directory conversion — recursively finds all supported files in a folder
- arXiv paper support — dedicated pdf2md.py tool with marker backend for complex multi-column PDFs
- Natural language triggers — “ingest this paper”, “ingest report.pdf” — no CLI syntax required
Knowledge Graph
- Two-pass graph build — deterministic wikilink extraction (EXTRACTED edges) + semantic inference (INFERRED/AMBIGUOUS edges with confidence scores)
- Louvain community detection — clusters related topics automatically
- Interactive graph.html — self-contained vis.js visualization, opens in any browser
- SHA256 cache — only changed pages are reprocessed
Quality & Governance
- Contradiction flags — new sources contradicting existing claims flagged at ingest time
- Lint reports — orphan pages, broken links, missing entity pages, data gaps with suggested sources
- Living overview — wiki/overview.md revised on every ingest to reflect current synthesis
Integration
- Obsidian integration — wiki designed to be browsed in Obsidian with consistent wikilinks
- Slash commands (Claude Code) — /wiki-ingest, /wiki-query, /wiki-lint, /wiki-graph
- Query answers as wiki pages — explorations compound like ingested sources
- Entity + concept auto-creation — pages created for every person, company, project, idea mentioned
Data & Storage
- Persistent wiki — structured markdown pages that accumulate across sessions
- Graph JSON — persistent node/edge data with SHA256 caching
- No API key required — runs entirely with local agent (Claude Code, Codex, etc.)
Relationships
- llm-wiki — implemented-by: agent-skill implementation
- claude-skills — compatible-with: sibling skill form factor
- model-context-protocol — operates-with: agent skills can call MCP tools
- infmat-karpathy-llm-wiki — sibling-of: sibling skill package
Implications
llm-wiki-agent’s multi-model support matters for the user’s cross-tool goal (the vault is model-agnostic). It shows the pattern running under several LLMs, not just one. The query modelo-ideal-para-ingest synthesizes what the vault knows about model choice for ingest — the agent is a primary source for the “multi-model works” conclusion.
External Links
- Repository: https://github.com/SamurAIGPT/llm-wiki-agent