yjid/kensaku

Kensaku is a local-first desktop search launcher for Windows, macOS and Linux. Instead of only matching file names, it reads inside documents, code, logs and screenshots, then surfaces hits ranked by relevance, recency or size and opens them from a keystroke-summoned window. It fits people who remember content but not file names, and who want search to stay on their own machine.

What changes in the workflow

The main shift is searching by what you remember. A PDF saved as notes-final-2.pdf can be found by its text, and a screenshot with no useful name can be found by the error message visible in its pixels. Queries can be typed naturally ("big videos from last month") or with a query language for filters such as type, extension, folder, modified time and regex. Results update as you type, and Space previews PDFs, images and text.

Indexing is one pass: each file feeds both a name index and a text index, and changed files are picked up incrementally with renames and deletes tracked. The author reports name search over 100,000 files averaging 1.3 ms on a warm in-memory index, with about 12 MB for names and metadata; see the performance section and benchmarks for the exact conditions and comparison caveats.

Ways to use it

• Desktop app: the primary path, shipping for Windows, macOS and Linux. Builds are not yet code-signed, so both macOS and Windows show a first-run warning; the install guide covers the steps. The auto-updater is not wired up yet.
• Command line: kensaku-cli ships with each release and exposes text and JSON search output plus index state.
• AI agents over MCP: an MCP-capable client can point at kensaku-cli mcp; clients needing HTTP can use kensaku-cli serve, which binds loopback only and requires a bearer token. Agent tools are read-only and reindex is off unless enabled. Setup is described in the MCP section and the agent guide.

Requirements and limits

Users choose which folders are indexed and can exclude paths; hidden files and common build folders are skipped by default, and credential files such as .pem, .key and .pfx are excluded. OCR is included: PP-OCRv6 is the default with a roughly 6 MB model pack downloaded on first run, while Intel Macs fall back to Tesseract. JPEG 2000, fax and JBIG2 scans are not decoded yet. Thumbnails cover images only, and visual search is a beta download not yet part of results.

The author states Windows receives the most real-world testing, while macOS and Linux builds come from CI and would benefit from more hands-on use. Building from source needs Rust stable, Bun 1.1 or later, and platform webview libraries (or the included Makefile tasks). The index itself is sensitive: it holds file names and extracted text, so the author advises against syncing it to a shared location. The full threat model is in SECURITY.md.

The project is MIT-licensed. HotRepo recorded 197 total stars on its rust NEW-REPOSITORY board on 2026-10-29; that is a single dated snapshot, not a growth trend.