nvk/llm-wiki

LLM-compiled knowledge bases for any AI agent. Parallel multi-agent research, thesis-driven investigation, source ingestion, wiki compilation, querying, and artifact generation.

llm‑wiki – LLM‑compiled knowledge bases for AI agents

What it is – llm‑wiki is an open‑source system that lets large‑language‑model agents (Claude, OpenAI Codex, OpenCode, Pi, etc.) build, maintain, and query a personal or team‑wide wiki. The wiki stores Ideas, Research, Projects and Artifacts in a structured, Obsidian‑compatible folder layout, and the agents can automatically ingest sources, run parallel research, audit content, and promote vetted briefs into deliverable projects.

Key concepts

  • Sessions – optional per‑agent memory that records what the agent has done in a wiki session; can be disabled per user.
  • Adapters – a private, manifest‑driven plug‑in system that routes ingestion and write operations to specific providers (e.g., web, CSV, media‑wiki) while keeping authentication and side‑effects under strict control.
  • Research modes – one‑shot queries, deep multi‑agent research, thesis‑driven investigations, and specialist‑method selection.
  • Export checkpoints – docs/knowledge/<slug>/ snapshots that contain a deterministic, privacy‑minimized hand‑off of a project’s knowledge.

Supported runtimes

Runtime How you invoke it Read‑only query size Full research size
Claude Code (native) /wiki:… commands or @wiki skill wiki:query (≈2.8 KB) @wiki (≈11.7 KB)
OpenAI Codex (marketplace) $wiki-query or @wiki after installing the plugin same as Claude same as Claude
OpenCode (instruction file) Add the SKILL.md URL to opencode.json wiki-query/SKILL.md wiki-manager/SKILL.md
Pi / DS4 (local models) ./scripts/pi-wiki-query (read‑only) or pi --skill …/wiki-manager/SKILL.md (full)
Any other LLM agent Copy profiles/query-lite/SKILL.md (read‑only) or AGENTS.md (full)

Installation shortcuts

  • Claude Code – claude plugin install wiki@llm-wiki
  • Codex – codex plugin marketplace add nvk/llm-wiki && codex plugin add wiki@llm-wiki
  • OpenCode – add the remote SKILL.md URL to opencode.json (or curl it locally)
  • Pi – pi --skill path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md

Typical workflow

  1. Create / select a topic – /wiki:research "nutrition" --new-topic
  2. Ingest sources – /wiki:ingest https://example.com/article or bulk import with ingest‑collection.
  3. Run research – @wiki research "gut‑brain axis" --deep --min-time 2h (spawns multiple agents that read, summarize, and cross‑reference).
  4. Query the compiled knowledge – /wiki:query "How does fiber affect mood?" (read‑only) or wiki-query from any LLM.
  5. Promote a brief – @wiki audit --project coldcard-threat-model moves a vetted brief into a delivery project.
  6. Export a checkpoint – v0.24.0 adds a command that writes a reproducible snapshot under docs/knowledge/ for hand‑off.

Architecture highlights

  • Claude‑first source of truth – the canonical skill lives in claude-plugin/skills/wiki-manager/SKILL.md. Runtime‑specific wrappers (Codex, OpenCode, Pi) are generated automatically via sync scripts.
  • Private‑adapter protocol (llm‑wiki‑adapter/v1) – adapters declare intent, URL routes, and read/write permissions. No provider credentials are ever baked into the wiki; the adapter enforces exact‑allowlists and idempotency keys.
  • File‑system hub – the wiki lives in a user‑configurable directory (~/.config/llm-wiki by default, fallback ~/wiki). The system is fully Obsidian‑compatible, so users can edit notes manually if desired.
  • Sandbox friendliness – explicit nono sandbox policies are documented for each runtime, ensuring the agent has the minimal read/write access it needs.

Why it matters

  • Gives LLM agents a persistent, searchable knowledge store instead of a stateless prompt.
  • Enables reproducible research pipelines: ingest → parallel analysis → audit → export.
  • Works across the major LLM‑agent ecosystems (Claude, Codex, OpenCode, local models) with a single source of truth.

Limitations / cautions

  • The system assumes the host agent can read/write the wiki directory; sandboxed environments must grant those permissions.
  • Full‑research commands are write‑capable and will modify files; use the read‑only wiki‑query skill when you only need look‑ups.
  • Exported checkpoints are read‑only; they never automatically commit or publish changes.

Where to learn more

Related

  • Project
    nashsu/llm_wiki

    A desktop application that automatically transforms documents into a structured, interlinked personal wiki using LLMs, featuring a knowledge graph and automated deep research.

  • Project
    Ar9av/obsidian-wiki

    A tool that allows AI agents to build and maintain a persistent, interconnected markdown knowledge base (a "second brain") from chat logs and documents.

  • Project
    atomicstrata/llm-wiki-compiler

    A compiler that turns raw sources into a structured, citation-traceable Markdown wiki, moving knowledge extraction from query-time to compile-time for better agent and human retrieval.

  • Project
    zosmaai/pi-llm-wiki

    A self-maintaining, Obsidian-compatible knowledge base for AI agents that transforms raw sources into a durable, interlinked wiki using the LLM Wiki pattern.

  • Project
    SamurAIGPT/llm-wiki-agent

    A coding agent skill that converts source documents into a persistent, interlinked markdown wiki with automated entity extraction and contradiction flagging.