Your research folder, ghostwritten by Claude

What it does

Point LLM Wiki at a folder full of PDFs, spreadsheets, and notes. It indexes everything into a local SQLite database, then exposes the contents to any MCP-capable agent—Claude, Cursor, or a local Llama model—via a stdio server. The agent reads your sources and writes ordinary markdown pages into a wiki/ directory, keeping cross-references and citations in sync while your original files stay completely untouched.

The interesting bit

The value is in the accumulating structure, not the chat. Instead of re-deriving summaries from scratch every session, the agent incrementally updates entity pages, fixes stale links, and flags contradictions across the workspace. The README nails the diagnosis: personal wikis die from maintenance fatigue, not lack of intent. This offloads the bookkeeping while leaving you in charge of the thinking.

Key highlights

• Runs fully offline by default; no cloud services or API keys needed for basic indexing and writing
• Filesystem is the source of truth: wiki/ pages are plain markdown editable anywhere, and the SQLite index is disposable and rebuildable
• Works with any MCP client, not just Claude—opencode, Continue.dev, Cursor, or local models via Ollama
• Extracts text locally from PDFs, HTML, Excel, and more; optional Mistral OCR for table-heavy documents
• Self-hostable multi-tenant version available with Postgres, Supabase auth, and S3

Caveats

• Local search is SQLite FTS5 keyword search only—no semantic or vector search unless you switch to the hosted Postgres stack
• Default PDF table extraction is rough; prose works fine, but tables come through messy without the optional Mistral OCR upgrade
• No cross-process write locks, so concurrent edits from multiple MCP clients can overwrite each other

Verdict

Worth a look if your research folder is outgrowing your ability to maintain summaries and cross-references by hand. Look elsewhere if you need semantic search out of the box or a polished multi-user editing experience.