llm-wiki-cli

Agent-native persistent knowledge management CLI — compile knowledge once, query forever.

LLM Wiki is a CLI tool that creates and manages an AI-maintained knowledge base using Obsidian-compatible Markdown files. It provides search, graph analysis, sync tracking, and a skill system for AI coding agents (Claude Code, Codex) to autonomously ingest, query, and maintain your wiki.

Based on jackwener/llm-wiki.

Install

npm install -g @jupiterthewarlock/llm-wiki

Requires Node.js >= 20.

Quick Start

mkdir my-wiki && cd my-wiki
llm-wiki init

This creates the vault structure:

my-wiki/
  wiki/                # AI-maintained wiki pages (Obsidian-compatible)
  sources/             # Raw source documents, date-partitioned
  .llm-wiki/           # Config and sync state
  .claude/skills/      # Skills for Claude Code
  .agents/skills/      # Skills for Codex/OpenAI
  wiki-purpose.md      # Wiki scope and audience definition
  wiki-schema.md       # Page types, naming conventions, frontmatter schema
  wiki-agent.md        # Agent identity and ingest rules
  wiki-log.md          # Append-only operation log
  CLAUDE.md            # Claude Code bootstrap instructions
  AGENTS.md            # Codex/OpenAI bootstrap instructions

CLI Reference

llm-wiki init [directory]

Initialize a new wiki vault.

| Option | Description | |--------|-------------| | [directory] | Directory to initialize (default: .) | | --existing | Adopt an existing directory without overwriting files | | --language <lang> | Vault language: en or zh (default: en) |

llm-wiki search <query>

Search wiki pages using BM25 keyword search. When DB9 is configured, also runs vector similarity search and merges results via Reciprocal Rank Fusion (RRF).

| Option | Description | |--------|-------------| | -n, --limit <number> | Max results (default: 10) | | --bm25-only | Skip DB9 vector search even if configured |

Output includes slug, title, description, snippet, score, and tags for each result. Supports English and CJK (Chinese/Japanese/Korean) tokenization.

llm-wiki graph

Analyze the wikilink graph. Detects communities (label propagation), hub pages, orphan pages (no incoming links), and wanted pages (linked but not yet created).

| Option | Description | |--------|-------------| | --json | Output analysis as JSON |

llm-wiki status

Show vault statistics and health summary: page count, source count, link count, log entries, last sync time, recently modified pages, and health checks (missing files, broken links, pages without sources, legacy filename detection).

llm-wiki sync

Track file changes and update sync state using mtime + SHA-256 content hash. Reports added, modified, deleted, and unchanged files. Auto-appends a sync summary to wiki-log.md. When DB9 is configured, upserts/deletes embeddings for changed pages.

| Option | Description | |--------|-------------| | --dry-run | Show changes without updating state |

llm-wiki log <verb> <subject>

Append a structured entry to wiki-log.md.

| Option | Description | |--------|-------------| | -d, --detail <text> | Additional detail line (repeatable) | | --max <n> | Max log entries to keep (default: 200) |

llm-wiki log ingest "readme-design-patterns.md" -d "created 3 wiki pages" -d "added 12 wikilinks"

llm-wiki skill

Manage AI agent skills.

| Subcommand | Description | |------------|-------------| | install | Install/upgrade skills to the workspace | | show <name> | Print raw skill markdown to stdout | | list | List all available skills |

| Install Option | Description | |----------------|-------------| | --claude | Install to .claude/skills/ only | | --codex | Install to .agents/skills/ only | | --dir <path> | Workspace directory (default: cwd) |

Default (no flags) installs to both .claude/skills/ and .agents/skills/.

Agent Skills

The bundled llm-wiki skill provides six AI agent operations:

| Skill | Description | |-------|-------------| | /ingest <path> | Read a source, extract entities/relationships, create wiki pages with [[wikilinks]] | | /query <question> | Search wiki, synthesize answers, optionally write back novel synthesis | | /lint | Health-check sweep for structural, content, and source issues | | /research <topic> | Deep investigation with external sources | | /enrich <file> | Add [[wikilinks]] to a wiki page by matching against the wiki index | | /graph | Interpret the knowledge graph: communities, hubs, gaps |

Configuration

Config is stored at .llm-wiki/config.toml:

[vault]
name = "My Wiki"
language = "en"    # "en" or "zh"

# Optional: DB9 vector search integration
# [db9]
# url = "postgresql://..."

The CLI walks up from the current directory to find .llm-wiki/config.toml and locate the vault root. Alternatively, set the LLMWIKI_VAULT_PATH environment variable to the vault root path to skip directory walking.

Wiki Page Format

Each page is a Markdown file with YAML frontmatter:

---
title: Page Title
description: One-line summary
aliases: [alternate names]
tags: [domain tags]
sources: [2026-04-22/source-file.md]
created: 2026-04-22
updated: 2026-04-22
---

Page content with [[wikilinks]] to other pages.

Pages support [[target]] and [[target|display text]] wikilink syntax, compatible with Obsidian.

DB9 Vector Search (Optional)

When a DB9 (PostgreSQL + pgvector) connection is configured, search uses hybrid ranking:

• BM25 for keyword matching (always available)
• Vector similarity via DB9's server-side embedding() function (1024-dim)
• Reciprocal Rank Fusion (K=60) to merge both result sets

License

Apache-2.0 — see LICENSE.