LLMwiki

LLMwiki turns a team's raw research folder into a reviewed, queryable, self-improving Obsidian wiki. Headless-first. Cites, doesn't guess.

Language: English (this page) · 简体中文 — Guide: English · 简体中文 — Wiki: Home · Architecture · Rationale · FAQ

You are reading this because your team has already lost knowledge.

Not because nobody wrote it down. They did: papers, meeting notes, repo findings, screenshots, agent answers. The problem is worse: the knowledge has no state. No source. No reviewer. No promotion path. No way to tell a draft from team truth.

LLMwiki gives that mess a compiler pass:

capture -> compile -> ask -> file -> review -> promote

Put source material in raw/. Compile it into wiki/ summaries, concept pages, backlinks, and contradiction reports. Ask agents cited questions. File useful answers into 00-Inbox/AI-Output/. Promote only reviewed knowledge into decisions, architecture, and runbooks.

It is not an AI companion. It is a reviewed team memory compiler. Obsidian is the IDE, Git/Gitea review is the ledger, and MCP/CLI tools are the execution surface.

Inspired by Andrej Karpathy's LLM Wiki. Markdown is the source of truth; the compiler turns structure into a graph; MCP exposes it.

Quick start (10 seconds, Claude Code)

Inside any Claude Code session:

/plugin marketplace add 2233admin/obsidian-llm-wiki
/plugin install llmwiki@obsidian-llm-wiki

That's it. No clone, no build, no config file to edit. The plugin ships the MCP server (runs from the plugin directory, Node 20+), all /llmwiki:vault-* knowledge-work roles, and the thinking/research commands. Start Claude Code inside your vault and the server finds it automatically (cwd is the vault); otherwise set VAULT_MIND_VAULT_PATH or drop a vault-mind.yaml.

Other hosts (Codex / OpenCode / Gemini)

git clone --depth 1 https://github.com/2233admin/obsidian-llm-wiki.git
cd obsidian-llm-wiki && ./setup                      # --host codex | opencode | gemini

Windows: .\setup.ps1. The script copies the skill bundle into your host's skills directory and prints the .mcp.json snippet to paste into your agent config. docs/INSTALL.md has per-host paths and the manual recipe.

See the loop (5 minutes)

You can verify the compiler loop before wiring any agent host. This demo is local, report-only, and the compiler dry-run uses stub extraction, so it does not need an API key.

python compiler/compile.py examples/collab-vault/research-compiler --tier haiku --dry-run
python scripts/knowledge_health.py --vault examples/collab-vault --json
python scripts/llmwiki_doctor.py --vault examples/collab-vault --json

Then inspect the before/after:

That is the product: raw material becomes cited, inspectable, reviewable team memory.

Works with

Any MCP-compatible host:

Anything else speaking stdio MCP transport should work — the setup script only copies skills into the right directory and prints the .mcp.json snippet. If your host reads MCP config from somewhere else, paste the snippet there by hand.

Example prompts

Cold start -- no vault context:

/vault-librarian what do I know about attention heads

Warm start -- specify a note you have:

/vault-librarian explain [[retrieval-augmented-generation]] in the context of my other notes on LLMs

Format-specific -- you want a list, not prose:

/vault-historian what decisions did I make about training data between January and March 2026

Iterate -- refine an answer:

/vault-curator find all orphan notes and stale notes in my vault that have not been updated in 90 days

Compile, Query, Govern

See docs/RESEARCH_COMPILER_LOOP.md for the standard operating loop.

Knowledge roles, one MCP surface

Each /vault-* command is a knowledge-work role over the same 40-operation MCP tool set. They are jobs in the pipeline, not product mascots.

Structured notes (v2.4.0)

Six tools that create AI-First notes with full frontmatter, wikilinks, and a "For future Claude" preamble — safe by default (dryRun: true).

Every note gets ai-first: true in frontmatter and a two-sentence preamble so a future Claude can decide relevance in seconds without reading the full note.

vault.init (v2.5.0) scaffolds a methodology layout — generic, para, lyt, or zettelkasten — into an empty or existing vault, safe by default (dryRun: true). All write operations are now covered by per-file advisory locking with a 60s TTL, so multiple agents can write to the same vault concurrently without clobbering each other. Inspired by claude-obsidian.

Thinking and research commands (v2.4.0)

Thirteen slash commands in commands/ for use in any Claude Code, Codex CLI, Gemini CLI, or OpenCode session. They run reasoning over the vault using the MCP tools above — no LLM logic lives in the server.

Inspired by obsidian-second-brain and claude-obsidian. vault-mind provides the infrastructure; these commands provide the workflow patterns that sit on top.

How it works (30-second tour)

Your markdown files -- with wikilinks [[like this]], aliases, frontmatter tags, and mtime -- are the source of truth. The compiler turns raw topic folders into a concept graph (nodes = notes, edges = links and semantic relationships), summaries, and concept pages. The MCP server exposes this graph as tools: vault.search, vault.backlinks, vault.graph, and 40+ more.

When Claude Code (or any MCP-compatible agent) runs /vault-librarian, it calls vault.search and vault.read directly. The agent gets citations -- not guesses.

• No embeddings required at small scale. Optional pgvector-backed semantic search via the memU adapter.
• No database. Filesystem-only by default; a compiled graph is cached as plain JSON alongside the vault.
• No Obsidian required at runtime. The filesystem adapter is always available. Obsidian is an optional adapter if you want live plugin-API features via a WebSocket bridge.
• No code intelligence required at small scale. Optional project-wide knowledge graph (code + docs + PDFs + images) via the graphify adapter (uv tool install graphifyy).

Deep dives

The wiki has the long-form answers. Read them in any order.

License

GPL-3.0. See LICENSE.