What it does

LLM Wiki indexes a folder of your existing research — PDFs, spreadsheets, markdown, whatever — and exposes it to Claude (or any MCP-capable client) as a structured workspace. Claude can search sources, write wiki pages, edit them, and maintain cross-references and citations. The wiki lives as ordinary markdown files on your disk; the SQLite index is purely a derived search layer you can rebuild anytime.

The interesting bit

The insight is maintenance, not generation. Most personal wikis die because someone has to update the 15 pages affected by one new paper. Here the LLM does that bookkeeping — re-deriving summaries, fixing stale links, flagging contradictions — while you stay in the loop on source selection and analysis. The filesystem remains the source of truth; the agent is just an unusually diligent editor.

Key highlights

• Local-first: indexing, storage, and file writes run on your machine; no API keys needed for basic usage
• MCP-native: works with Claude Desktop/Code, Cursor, Cline, Continue.dev, opencode, or any MCP client
• One-command start (./llmwiki open ~/research) scaffolds the workspace, starts the Next.js/FastAPI stack, and opens the browser
• Document pipeline handles PDFs (pdf-oxide, with optional Mistral OCR for tables), HTML, Excel/CSV, and office files via optional LibreOffice
• Self-hostable multi-tenant version with Postgres, Supabase auth, and S3 (powers llmwiki.app)

Caveats

• PDF table extraction via pdf-oxide is rough; data-heavy documents need Mistral OCR for usable output
• No vector/semantic search in local mode — SQLite FTS5 only; hosted version uses PGroonga
• No cross-process write locks: concurrent edits from multiple clients can clobber each other
• LibreOffice required for Word/PowerPoint extraction; skip it if you live in PDFs and markdown

Verdict

Researchers, analysts, or anyone whose ~/Downloads has become a graveyard of half-read papers should try this. If you already have a rigid Zettelkasten that you maintain obsessively by hand, the automation may feel like overkill — or like admitting defeat.