Feature request: attach images to annotations (local temp storage + cleanup on close)

[mehdis34]

Hi! I’m using Agentation with MCP (Codex) and would love a way to include visual context in annotations.

Request :
Add support for attaching/uploading one or more images directly inside an annotation (screenshots, mockups, reference images), so the agent can use them as context.

Suggested approach (local-first)
For the MCP use case, it would be awesome if attachments could be stored locally on the machine running the app/agent (temporary folder), and then deleted automatically when the annotation is closed/resolved.
Store images in a temp directory (e.g. .agentation/tmp or OS temp) with stable IDs.

Annotation payload includes only lightweight metadata + a local file reference, e.g.:

{
  "attachments": [
    {
      "type": "image",
      "id": "att_123",
      "filename": "screenshot.png",
      "mimeType": "image/png",
      "localPath": "/…/agentation/tmp/att_123.png"
    }
  ]
}

When the annotation is marked as closed/resolved, Agentation deletes associated files (and optionally a GC job cleans old orphaned files).

UX ideas

• "Attach image" button in the annotation composer
• Drag & drop + file picker
• Paste screenshot from clipboard would be a big win

MCP / Webhooks
MCP tools should expose attachments metadata (and ideally allow reading the file contents or serving them via a local endpoint).
Webhooks could include either URLs (if hosted) or omit binary data and only send metadata (depending on design).

Why local temp storage?
Keeps images private by default (no external hosting required)
Lightweight payload (avoid base64 bloat)
Works well for agent workflows on a dev machine
Auto-cleanup on close keeps disk usage under control
Happy to help test or clarify edge cases (Windows/macOS/Linux paths, security, limits, etc.).

[prakersh]

You might also look at onUI — it's a free, open-source browser extension that captures annotations with element metadata (CSS selectors, page URL, viewport info) and exposes them to AI agents via a local MCP server. It takes a metadata-rich approach to annotation context rather than image attachments.