screenpipe-search_skill

Overview

This skill lets you search your locally recorded screen text, audio transcriptions, and UI events through the Screenpipe API running at http://localhost:3030. It helps you find what you saw, heard, or clicked on by querying OCR, audio, vision frames, and input events. All searches run locally against your personal recordings and respect strict time-range requirements to avoid timeouts.

How this skill works

The skill issues HTTP requests to the Screenpipe search endpoint, returning OCR text, audio transcripts, and UI event records. Queries must always include a start_time and should start with a narrow window (last 30–120 minutes). When needed, it fetches specific screenshot frames via frame_id, but prefers OCR/text first to conserve tokens.

When to use it

• When you want to find text or images that appeared on your screen within a time range.
• To locate what was said in a meeting or who spoke using audio transcriptions.
• To audit recent UI interactions: clicks, keystrokes, app switches, or clipboard events.
• When you need to reconstruct activity across apps (e.g., browser + Slack + code editor).
• When you want a screenshot for visual confirmation after OCR results are ambiguous.

Best practices

• Always include start_time in ISO 8601 format; default to the last 1–2 hours and expand only if needed.
• Use app_name or window_name filters when the user mentions a specific app to narrow results and improve speed.
• Keep limit low (5–10) on initial queries to avoid large responses and timeouts.
• Treat “recent” as last 30 minutes, “today” as since midnight, and “yesterday” as the previous date range.
• Fetch frames only when OCR or transcripts are insufficient; never fetch more than 2–3 frames per query due to token costs.

Example use cases

• Find what you discussed in this morning’s 1-hour meeting by searching audio with start_time set to the meeting start.
• Locate a browser tab’s text from two hours ago by searching OCR with app_name=Google Chrome and a 2-hour window.
• Show what you clicked in a form by querying input events for a focused window and the relevant time range.
• Confirm the content of a dashboard chart by searching OCR for keywords, then fetching one key frame for visual context.
• List everything you worked on today by running short-range searches per app (Chrome, Slack, Code) and summarizing results.