A desktop knowledge layer that leaves your files alone

What it does

StashBase is an Electron desktop app that opens a local folder and overlays it with a source-linked Markdown wiki and a rebuildable semantic index. It extracts text from PDFs, DOCX files, images, and even local audio or video transcriptions, then exposes the results to a built-in Claude Code or Codex chat and to external agents through a local MCP server. Your originals stay untouched on disk.

The interesting bit

Instead of importing files into a proprietary silo, it treats the folder as the source of truth, adding only a wiki/ directory and rebuildable index metadata. Even audio transcription runs offline via downloadable local speech models, so sensitive recordings can stay on the machine unless you explicitly opt into cloud embeddings.

Key highlights

• Indexes PDF, DOCX, images, and media via local OCR and transcription, leaving originals in place.
• Serves the library to external AI clients through a local MCP server while the app is running.
• Exact-text search works out of the box; semantic search requires an OpenAI, OpenRouter, or StashBase-hosted embedding key.
• Built-in chat uses Claude Code or Codex runtimes when available, prompting for installation if they are missing.
• Ships as a signed desktop app for macOS Apple Silicon and Windows x64, with a community-supported Linux build.

Caveats

• The project is explicitly labeled early alpha.
• Linux support is community-supported and limited to x86_64 Debian 12+ and Ubuntu 22.04+.
• Audio transcription and semantic indexing require a separately downloaded local model or an external API key.

Verdict Worth a look if you want a local-first, agent-friendly layer over an existing folder of notes and documents. Skip it if you need a mature, fully cross-platform solution or a server-side web app.