Inginnng/EditHere

What it is

EditHere („改这里“ — change here in Chinese) is a desktop tool that turns vague UI feedback into precise visual instructions for AI. Instead of typing „move the button a bit to the left“, you circle the button, drag it, or pause a video at the right frame — and EditHere packs everything into a structured file an AI can act on.

It runs locally on Windows, macOS, and Linux, and it never calls an AI model itself. It just prepares the feedback; any AI tool (Codex, Claude Code, and others) can consume it.

The problem it solves

When you ask an AI to change an interface, describing where and how much with words is slow and error-prone. EditHere removes the guesswork:

• Where — you click the exact spot instead of describing it.
• How much — you drag/resize the element; EditHere records the real before/after coordinates.
• Multiple edits — each comment is numbered and matched to a screen position.
• Video/animation issues — pause at the exact frame, comment, and the export includes a timestamp and that frame's image.

How it works

1. Capture — press the global hotkey (Alt + Shift + 2) or ask the AI to start. Hovering identifies interface elements; the wheel zooms the selection; you can also drag a box manually.
2. Comment — use a point, a box, a global note, or the „Explode“ mode (see below). Write your thoughts right on the image.
3. Export — one click produces a JSON file. It contains the original image, every annotation with coordinates, and any element movements.
4. Hand off to AI — either copy the JSON manually into any chatbot, or (with the bundled edithere-cli and skill) the AI calls edithere-cli annotate, waits, and reads the feedback the moment you click „Done and return to AI“. Cancelling or timing out does not submit uncommitted edits.

Key features (from the README)

• Smart screenshot — hover to catch a UI element, manual box selection, magnifier with pixel coordinates and color readouts (RGB/HEX/HSV/HSL), fixed aspect ratios (1:1, 4:3, 16:9, 9:16, custom), and history of the last 20 captures.
• OCR (text recognition) — selected area is read line by line; clicking a line highlights it on the image. Uses the OS's own engine (Windows OCR / macOS Vision) — offline, no uploads, no new dependencies.
• Long screenshots — select a scrolling region and the tool scrolls and stitches a tall image, with a live thumbnail. Duplicate fixed header is kept only once.
• Annotations — point, box, or whole-screen comments, each numbered and mapped to the image. One click exports JSON.
• „Explode“ mode — splits the screen into movable components. You drag, scale, or type exact coordinates; each move becomes an annotation recording the before/after area. The original spot is left blank so the AI knows something must fill it.
• Video annotation — pause at any frame, comment, then jump to another time point. Exported JSON includes timestamps and frame captures.
• Project save — .edithere files store the original image, annotations, and edit state, so you can reopen and continue later.
• Inputs — PNG / JPEG / WebP / BMP images and MP4 / MOV / WebM videos, via screenshot, open dialog, drag-and-drop, or paste.
• Local-first — screenshots, element detection, OCR, and editing all happen on your machine. Update checks contact GitHub/Gitee only for version info (can be turned off).

JSON example (from the README)

The exported feedback looks like this (trimmed):

{
  "image": "data:image/png;base64,…",                     // original image (optional)
  "annotationSpace": "result",
  "objects": [
    { "source": {"x1":1002,"y1":156,"x2":1464,"y2":384},
      "movements": [],
      "annotations": ["Make the conversion rate drop more prominent, red label"] },
    { "source": {"x1":1760,"y1":50,"x2":1960,"y2":110},
      "movements": [{"to": {"x1":640,"y1":40,"x2":840,"y2":100}}],
      "annotations": ["Move the export button next to the title"] },
    { "source": null, "movements": [], "annotations": ["Add more whitespace overall"] }
  ]
}

Platforms & install

• Windows 10 1809+ — officially supported. Installer EXE (no admin rights needed) or portable ZIP. Not code-signed yet; checksums available.
• macOS 14+ — preview stage, not yet notarized. Universal DMG.
• Linux x86_64 — preview, targeted at Ubuntu 24.04. AppImage.
• The README also includes prompts that let an AI tool (Codex, Claude Code) do the whole installation for you.

License — important

PolyForm Noncommercial 1.0.0 — free for personal, educational, research, charitable, or governmental use (modify and redistribute allowed with license text). Commercial use requires written permission from the author (contact inginnng@163.com). The name and logo are not covered by the license.

Docs & contact

Docs cover user guide, CLI/agent integration, AI-assisted install, video annotation format, Linux notes, build/development, and changelog. Feedback via GitHub issues or QQ group 1018416966.

---

All information above is taken directly from the project README; no features have been added or invented.