A free, open-source local file search launcher: search file names, file contents (PDF, DOCX, EPUB, etc.), text in screenshots (OCR), and semantic meaning all at once. It also allows agents like Claude and Cursor to perform read-only searches on the same index via MCP.
Stars 221 (+216 / 24h)
Reasons
- The author claims that on Linux, searching file names across 10 ten-thousand files averages 1.3 milliseconds, with file names and metadata consuming about 12 MB of memory; indexing 101101 files for the first time took 184 seconds
- Supports natural language time expressions ("screenshots from last week") and query syntax (type:pdf, size:>10mb, modified:7d, re: regex, etc.); OCR defaults to PP-OCRv6, falling back to Tesseract on Intel Macs
- MCP is read-only by default, reindexing is disabled by default, returned content has size limits and is returned as data rather than instructions; no telemetry is collected, and credential files such as .pem and .key are not indexed by default
Usage
- Download the Windows exe / msi, macOS universal dmg, or Linux AppImage / deb / rpm from Releases; on first run, select the folders to index, with optional downloads for OCR and model packages
- In MCP clients, set command to kensaku-cli and args to ["mcp"]; for HTTP-based clients, use kensaku-cli serve (listens only on localhost, requires a token)
Risks
- Installers are not yet code-signed, so macOS and Windows will show warnings on first run; automatic updates are not yet connected
- The index contains file names and extracted text, and the author advises against syncing it to shared locations; performance figures are self-tested by the author on Linux and have not been compared with Spotlight or Everything; the repository was created on 10 24, licensed under MIT