A research department that lives inside your coding agent

What it does

LLM-Wiki is a knowledge-management protocol that runs inside AI agents like Claude Code and OpenAI Codex. It orchestrates multi-agent research workflows—ingesting sources, running thesis-driven investigations, and compiling the results into linked markdown wikis that are compatible with Obsidian. The output is a persistent, queryable knowledge base rather than a one-off chat answer.

The interesting bit

Instead of building a standalone app, the project treats the behavior as a portable skill: Claude Code is the primary user, and thin packaging layers are generated for Codex, OpenCode, Pi, and any generic agent via a single AGENTS.md file. This means the same research logic survives across competing LLM ecosystems without hand-maintained ports.

Key highlights

• Ships as native plugins for Claude Code and Codex, plus instruction files for OpenCode, Pi, or any LLM agent that can read AGENTS.md.
• Supports “collector” catalogs for provenance-rich datasets of media, memes, tools, and entities, with bounded media downloads and scale-aware handling.
• Topic archive lifecycle lets you retire old wikis to .archive/ without polluting active context.
• Obsidian-compatible output: the compiled wikis are plain markdown with frontmatter.
• Includes linting, audit, and repair tools for its own registry and article indexes.

Caveats

• iCloud and sandbox paths require explicit permission configuration; the README dedicates a section to nono sandbox and macOS privacy workarounds.
• The Codex, OpenCode, and Pi runtimes are generated mirrors of the Claude source of truth, so non-Claude users are essentially running a translated skill.
• Several recent patch releases focused on “lint repair correctness” and “safer lint defaults,” suggesting the internal bookkeeping can be finicky at scale.

Verdict

Worth a look if you want your AI assistant to do sustained, multi-source research and leave behind an organized wiki rather than a chat transcript. Skip it if you just need quick answers or don’t want to curate a long-term knowledge hub.