Open SourceMIT License

Snap

Visual evidence capture for technical analysis. Annotate any screen with structured metadata, then feed it directly to AI agents for investigation.

View on GitHub Request Implementation

CAPTURE_FLOW

01 Hotkey triggers overlay

02 Annotate what matters

03 Structured metadata saved

04 Agent reads it. Agent acts on it.

Structured Visual Evidence

Snap captures your screen, lets you annotate it with circles, arrows, rectangles, text labels, and numbered markers, then saves both the annotated image and structured JSON metadata to a local inbox.

An MCP server exposes the inbox to any AI agent. Every annotation carries coordinates, labels, colors, source window context, and display resolution. The agent does not interpret a screenshot. It reads structured evidence.

Where Snap Fits

Technical Diligence

Capture product UI, dashboards, and system states during evaluation. Structured annotations become evidence in the diligence record.

Operations Review

Document what operators actually see. Annotated screenshots surface gaps between what was designed and what is running.

Architecture Review

Mark up interfaces, data flows, and integration points. Feed annotated evidence directly to analysis agents.

Portfolio Monitoring

Capture product progress across portfolio companies. Visual evidence that compounds alongside intelligence pipelines.

Two Components

Capture Overlay

Tauri 2.x desktop app (Rust + JS). Global hotkey triggers a fullscreen overlay for annotation.

+Circle & rectangle tools

+Arrow & freehand drawing

+Text labels & numbered markers

+Color selection & stroke widths

+Undo support

+Dim toggle for complex screens

MCP Server

Python server exposes the annotation inbox via MCP. Any compatible agent reads your evidence with full structured metadata.

+check_new_annotations()

+get_latest_annotation()

+list_annotations(last_n)

+get_annotation(filename)

+clear_inbox()

Evidence Format

Every capture produces a matched pair: an annotated PNG and a structured JSON file with annotation coordinates, labels, colors, source window title, process ID, display resolution, and timestamp.

Agents use the structured data to understand exactly what you marked, where, and in what context. No ambiguity.

{
  "source": {
    "window_title": "localhost:3000/dashboard",
    "resolution": [3840, 2160]
  },
  "annotations": [
    {
      "type": "circle",
      "center": [340, 220],
      "label": "data mismatch"
    }
  ]
}

Works With

Claude Code

Claude Desktop

Cursor

Any MCP Client

Linux (X11/Wayland)

macOS

If you can see it, you can evidence it.

Open source. MIT licensed. Built by Adjective.

View on GitHub See All Capabilities