Your PC is the system of record
Screen frames, extracted text, search indexes, embeddings, settings, and local model files live under the app data directory on your Windows machine.
Windows 10/11 · Local-first · InspectablePrivacy is the
Privacy is the
architecture
ScreenSearch keeps capture, OCR, storage, search, and AI inference on your Windows PC. Model downloads and signed update checks can use the network without uploading your screen history. This page explains what data exists, where it lives, and what controls are available.
No ScreenSearch cloud account Capture starts only when enabled Source-linked claims
00 — The Short Version
Privacy gates run before reading
ScreenSearch can skip locked sessions, excluded apps or titles, its own windows, and frames that do not pass capture rules.
No keystroke or clipboard capture
The app does not rely on keyboard hooks, typed content, clipboard contents, or pointer input to build screen history.
Network use is narrow
AI inference runs locally. First-use model or binary downloads and signed release checks may use the network without sending captured context.
01 — What Exists On Your PC
The private thing is not abstract. It is files and rows.
ScreenSearch uses the app data directory as the local boundary. The exact Windows path depends on the installed app, so this page uses <app-data> for that root.
- D.01
SQLite database
Searchable local store
Frames metadata, extracted text, spans, search tables, vector indexes, jobs, settings, and reports live in screensearch.db.
- D.02
Frame images
Local screenshots
Changed frames are stored as lossless WebP files below frames/day-*/ until retention or deletion removes the image files.
- D.03
Models
Local AI files
Embedding and local answer/vision models are stored under local model directories after download.
- D.04
Sidecar
Local inference runtime
The llama.cpp sidecar is stored locally and serves inference over loopback when local answers are enabled.
- D.05
Logs
Operational diagnostics
Logs are local. Screen content and OCR text should not be written into info-level logs.
ScreenSearch does not need a ScreenSearch cloud account to build this history. The website also does not run capture, indexing, inference, or product storage in your browser.
02 — Capture Boundaries
The most important privacy control is what never enters the history.
ScreenSearch treats capture as a gated pipeline. Configured app, process, and window-title exclusions are checked case-insensitively before a frame is stored.
- 01
Capture starts disabled
The app initializes with capture off. Screen history starts only after the user enables capture.
- 02
Foreground and lock gates run first
Pause-on-lock, excluded apps, excluded window titles, and ScreenSearch's own windows can stop capture before text is read.
- 03
Sensitive apps can be excluded
Password managers, banking apps, private browsers, medical records, client documents, and confidential work tools belong on the exclusion list.
- 04
Input content is not the trigger
Event-aware capture uses foreground and idle signals; it does not store typed keys, clipboard contents, clicks, scrolls, or pointer input.
03 — Reading And Indexing
ScreenSearch reads what is useful, then makes it searchable locally.
The app reads accessible foreground-window text first and falls back to OCR when content only exists as pixels.
UI Automation first
UIA can read foreground-window accessibility text without re-OCRing every pixel. Password fields are skipped, offscreen or occluded elements are dropped, and spans outside the target window are filtered out.
OCR fallback for pixels
When useful text is only visible in an image, PDF, terminal, browser surface, or older app, native OCR turns visible pixels into local text spans for search.
Attention-first filtering
The store separates raw extracted text from more useful content text, so app chrome and repeated surfaces do not have to dominate recall.
Local search indexes
Keyword search, vector search, and retrieval metadata are built from local rows. Search answers are grounded in source frames you can inspect.
04 — AI And Network Boundaries
AI inference stays local. Network access has narrower roles.
Current releases run AI through a bundled llama.cpp sidecar bound to the local machine. Network access supports model downloads and signed release checks, not remote inference over your captured context.
- A.01
Embeddings
Text embeddings are generated locally after the model is available on the machine.
- A.02
Local sidecar
Answer and vision workflows can run through a local llama.cpp sidecar bound to loopback, such as 127.0.0.1 on an ephemeral port.
- A.03
Downloads
First-use model or binary setup may contact hosts such as GitHub or Hugging Face. These downloads do not upload your screen history.
- A.04
Updates
The signed GitHub Releases update check uses the network to discover and download app releases without sending captured frames, OCR text, or search indexes.
05 — Retention And Deletion
Retention is a product behavior, not fine print.
Screen history can be useful and still deserve a short leash. The current app has separate behavior for screenshot images and searchable text.
Default image window
Current builds default to retaining screenshot image files for 30 days. A retention value of 0 keeps image files until you delete them.
Degrade to text
When image retention expires, the app removes the screenshot file and keeps searchable text, spans, and embeddings unless the frame itself is deleted.
User controls
Pause capture, exclude sensitive apps, and delete frames when history should not persist. For high-risk work, exclude first rather than cleaning up later.
06 — Inspectable Proof
Trust should have a trail.
These links point to the desktop app repository. Source evolves, so the page should be updated when privacy-relevant implementation details change.
- S.01
App data, startup, retention
src-tauri/src/lib.rs - S.02
Capture gates and own-window protection
crates/capture/src/lib.rs - S.03
Excluded app/title matching
crates/traits/src/privacy.rs - S.04
UIA password and offscreen filtering
crates/uia/src/lib.rs - S.05
Frame WebP storage pipeline
crates/kernel/src/capture_loop.rs - S.06
SQLite, FTS, vector schema
crates/store/src/schema.rs - S.07
As-built desktop architecture
docs/ARCHITECTURE.md
Install with eyes open.
A screen memory tool deserves scrutiny. ScreenSearch is designed so the answer to "where did my data go?" stays concrete.
No account Local-first Local AI inference