This project provides LLM-compiled knowledge bases, enabling AI agents to conduct parallel research, compile wikis, and generate artifacts. It targets AI researchers and power users seeking structured, queryable knowledge from diverse sources, offering automated thesis-driven investigation and artifact generation.

How It Works

The system employs a hub-and-spoke architecture where a central ~/wiki/ directory acts as a registry (wikis.json) for isolated topic sub-wikis. Each topic wiki maintains its own index, articles, and raw sources. Research is conducted via parallel agents (5 standard, 8 deep, 10 "retardmax" for aggressive ingestion), which ingest sources (URLs, files, text) into a topic's inbox/ and raw/ directories. Compiled articles are synthesized with cross-references and confidence scores. Queries can be performed at varying depths (quick, standard, deep), and artifacts like reports or slides can be generated. Key design principles include isolated topic wikis, dual-linking (Obsidian [[wikilinks]] and standard Markdown), immutable raw sources, and a structural guardian for integrity checks.

Quick Start & Requirements

Installation options include:

• Claude Code (native plugin): claude plugin install wiki@llm-wiki
• OpenAI Codex (repo-local plugin): Install via the repo's marketplace (.agents/plugins/marketplace.json).
• Portable AGENTS.md: Copy AGENTS.md into your agent's project root.

The project emphasizes zero dependencies when run with Claude Code.

Highlighted Details

• Parallel Research Modes: Offers standard (5 agents), deep (8 agents, adds historical/adjacent fields), and "retardmax" (10 agents, aggressive ingestion, skip planning) for varied research intensity.
• Thesis-Driven Research: Evaluates specific claims with dedicated agents (supporting, opposing, mechanistic), filters sources by relevance, and provides a verdict with anti-confirmation bias mechanisms.
• Obsidian Integration: Each topic wiki includes .obsidian/ configuration, allowing it to be opened as an independent Obsidian vault for graph visualization and backlinking.
• Dual-Linking: All cross-references use both Obsidian [[wikilinks]] and standard Markdown links, ensuring compatibility across Obsidian, Claude Code, and any text editor.

Maintenance & Community

The project is actively developed, with recent updates focusing on parallel research, Codex packaging, testing infrastructure, and project-aware linting. Community links such as Discord/Slack are not explicitly mentioned in the README.

Licensing & Compatibility

The software is released under the MIT License, which permits commercial use, modification, distribution, and sublicense, provided the original copyright and license notices are included.

Limitations & Caveats

While the AGENTS.md file offers portability, the core functionality is deeply integrated with Claude Code. Performance scaling beyond approximately 100 articles may require alternative local search engines like tobi/qmd. The "retardmax" mode prioritizes speed and breadth over immediate accuracy, requiring subsequent linting or refinement.