The launcher that reads your screenshots back to you

What it does

Kensaku (Japanese for “search”) is a Rust/Tauri launcher for Windows, macOS, and Linux that searches not just filenames but the text inside PDFs, Office docs, code, and images. It runs OCR on screenshots by default, so “that CUDA error I screenshotted” is a findable query. Everything stays on your machine — no account, no telemetry, no cloud.

The interesting bit

It exposes the local index to AI agents over MCP, read-only by default, with responses size-budgeted and file content returned as data rather than instructions — a rare bit of prompt-injection hygiene in agent tooling. The CJK-aware tokenizer (unigram + bigram for Chinese, Japanese, Korean) is also homegrown rather than borrowed.

Key highlights

• Filename search over ~100,000 files in ~1.3 ms average, ~12 MB of memory for the name index
• Full-text search across 50+ code/data formats plus PDF, DOCX, EPUB, and more
• OCR via bundled PP-OCRv6, with an optional higher-quality Indra engine (~27 MB) and Tesseract fallback
• Natural-language time queries (“screenshots from last week”) and a proper query language (type:pdf, re:"fn \w+_test")
• Ships a CLI (kensaku-cli) alongside the GUI; MCP over stdio or loopback HTTP with a bearer token

Caveats

• Builds aren’t code-signed, so expect OS warnings on first run; the auto-updater isn’t wired up yet
• Benchmarks are Linux-only, and the author admits the comparison against plocate/fd isn’t apples-to-apples (in-app fuzzy search vs. cold process starts)
• Windows gets the most real-world testing; macOS and Linux builds are CI-shipped and “would benefit from more hands”

Verdict

Worth a look if you’re on Windows and hoard screenshots, PDFs, and logs — it’s the rare content search that’s free, local, and agent-friendly. If you live in Spotlight or Raycast on macOS, the pitch is weaker until those builds get more mileage.