LockerGnomeFor serious tech professionals with a silly side.
Subscribe

Search

Softwareainotesopen-sourcelocal-firstobsidian

Wenlan Turns Your Notes and AI Chats Into a Local Wiki, and Throws Out Any Page Whose Citations Don't Match the Source

The open-source desktop app and CLI has the AI tool you already use draft cited wiki pages from your files and chat history, then refuses a page when most of its citations fail to match what they point at.

Chris pulls a card from a library card catalogue and drops it in a bin because its red thread doesn't reach the source book beside it.
Illustration: LockerGnome

The README’s one-line pitch is “Notes, code, AI chats into a wiki, checked against sources.” It reads Markdown, plain text, text-based PDFs, Word files and whole folders, plus Obsidian vaults and Logseq graphs. Web links get saved as sources. Gmail, Notion, Google Drive, Slack and GitHub come in only through your AI tool’s own connectors.

Chats are handled with some restraint. Sessions from Claude Code, Codex and others are saved as short memories: “Full transcripts are not imported,” and neither are attachments. Wenlan doesn’t bring its own model to write with. “The AI tool you already use turns your notes, files, and AI chats into pages, with no extra model.”

The refusal rule is the part worth stealing. Pages carry [n] markers, and each one opens the passage it claims to cite. When most of a page’s markers don’t line up with their sources, the page doesn’t make it into the wiki. Wenlan also flags pages that go stale when a source changes, and never edits your existing notes without your OK. The README is careful about what that buys you: “Passing these checks does not establish factual correctness.” It doesn’t say exactly where “most” begins.

Hosted note tools with an AI button will copy the gate before long, because a wiki that quietly invents footnotes is worse than no wiki at all. The local-first part is harder to copy, and it comes with a fine-print warning the README states outright: “A cloud AI client may send retrieved knowledge to its provider, and a cloud model receives the material it processes.” Local storage is only as private as the model you point at it.

The speed figure comes with an honest asterisk. On the maker’s own test, a 10,000-note Obsidian vault was “searchable in 15 s,” the first page appeared in about two minutes and eight cited pages in about four. That was version 0.20.2 on an M2 Pro Mac with 16 GB, a debug build, invented notes and Claude Code doing the writing: a single self-reported run. Everything here comes from the project’s own README.

The licensing splits in two. “The local runtime, CLI, MCP server, shared types, and plugins use Apache-2.0,” while the desktop app is AGPL-3.0. It runs on Apple Silicon Macs with macOS 11 or later and on Windows x64, and the Mac CLI installs with npx -y wenlan setup. With 89 stars on GitHub, nobody outside a small crowd has thrown a messy real vault at it yet.

More Software

Every Software story →