opencode-desktop-context

OpenCode plugin that captures desktop screenshots and adds them to the active session context.

I thought about Macrohard and whatever the Microsoft version was called and thought that was pretty cool — why not just make my own in OpenCode?

Beta warning: This plugin is early-stage software and may be unstable. Back up your OpenCode configuration before installing it, and consider having another coding assistant or terminal open so you can quickly identify and fix any issues. Pull requests and bug reports are welcome.

Privacy warning: Desktop screenshots can contain personal, sensitive, or confidential information. Anything captured may be sent to the model provider you have configured. Use this plugin at your own risk. Whenever possible, use a local model (for example, an Ollama model) so screenshots never leave your machine.

Installation

Install the plugin from npm and add it to your opencode.json:

npm install opencode-desktop-context

{
  "plugin": [
    ["opencode-desktop-context", {
      "captureTarget": "fullScreen",
      "maxAgeMs": 30000,
      "autoAttach": true,
      "systemHint": false,
      "visualIndicator": true,
      "retention": "temp",
      "retentionTtlMs": 600000,
      "blocklist": ["1Password", "Bitwarden", "Chase", "Keychain Access"],
      "allowlist": [],
      "quality": 80
    }]
  ]
}

Features

• Auto-attach: Adds the latest desktop screenshot to every user message.
• On-demand tool: Exposes capture_desktop for the model to request a fresh screenshot.
• Privacy controls: Permission persistence, blocklist/allowlist, configurable retention, and optional capture indicator.
• Cross-platform: Native capture adapters for macOS, Windows, and Linux.

Configuration

| Option | Type | Default | Description | |---|---|---|---| | captureTarget | "fullScreen" \| "activeWindow" \| "allDisplays" | "fullScreen" | What to capture. | | maxAgeMs | number | 30000 | Reuse a cached screenshot if newer than this. | | autoAttach | boolean | true | Attach latest screenshot to every user message. | | systemHint | boolean | false | Inject a hint telling the model about capture_desktop. | | visualIndicator | boolean | true | Show a brief indicator when a screenshot is taken. | | retention | "memory" \| "temp" \| "persistent" | "temp" | Where screenshots live. | | retentionTtlMs | number | 600000 | For temp, delete files older than this. | | persistentDir | string \| null | null | Required when retention is "persistent". | | blocklist | string[] | ["1Password", "Bitwarden", ...] | Window/app titles or names to never capture. | | allowlist | string[] | [] | If non-empty, only capture when active window matches. | | quality | number | 80 | JPEG quality for attachments (PNG if 100). | | visionModel | string \| null | null | Local Ollama vision model to describe screenshots (e.g. moondream:latest). When set, autoAttach sends a text summary instead of the raw image. | | ollamaBaseUrl | string | http://127.0.0.1:11434 | Base URL of your local Ollama server. | | periodicCaptureMs | number | 300000 | Automatically refresh the cached screenshot every N ms. Set to 0 to disable. |

Storage locations

• memory: the latest screenshot is kept in memory only.
• temp: screenshots are saved to ~/Pictures/opencode-desktop-context/ and cleaned up automatically based on retentionTtlMs.
• persistent: screenshots are saved to the directory you set via persistentDir.

Enabling auto-attach

autoAttach is true by default. When enabled, every user message you send is sent with the latest desktop screenshot attached. Change autoAttach to false if you only want the model to request a screenshot via the capture_desktop tool.

Privacy

Platform Requirements

• macOS: screencapture (built-in)
• Windows: PowerShell with .NET (built-in)
• Linux: One of grim (Wayland), import from ImageMagick, or ffmpeg (X11)

Vision-less Models

Some OpenCode models cannot process image attachments. If your model does not support vision, attach a local vision model to read screenshots for you.

Recommended: Moondream via Ollama

Moondream is a small, fast vision model that runs locally through Ollama. It is ideal for describing screens and reading text in screenshots.

Setup

1. Install Ollama if you do not already have it.

2. Pull the Moondream model:

   ollama pull moondream:latest

3. Send a captured screenshot to the model with the Ollama REST API:

   python3 - <<'PY'
   import base64, json, urllib.request
   
   image_path = "/tmp/opencode-desktop-context/capture-latest.jpeg"
   image_b64 = base64.b64encode(open(image_path, "rb").read()).decode()
   
   payload = {
       "model": "moondream:latest",
       "prompt": "Describe this screenshot and transcribe any readable text.",
       "images": [image_b64],
       "stream": False,
   }
   
   req = urllib.request.Request(
       "http://localhost:11434/api/generate",
       data=json.dumps(payload).encode(),
       headers={"Content-Type": "application/json"},
   )
   
   resp = urllib.request.urlopen(req)
   obj = json.loads(resp.read())
   print(obj["response"])
   PY

Tips

• Use the REST API (/api/generate) rather than ollama run < image.jpg; the CLI can interleave terminal control sequences that corrupt the output.
• For text-heavy screens, OCR with Tesseract is usually more accurate than a vision model.
• Moondream is the recommended default because it is small (~1.6B parameters) and fast. If you need richer descriptions, llava:latest or llava:13b are stronger but heavier alternatives.

Privacy

The first time capture_desktop is called, the plugin records a permission grant in ~/.config/opencode/desktop-context-permission.json. You can revoke permission by deleting that file. The blocklist is checked against the active window title and application name before every capture.

Local vision mode

To keep screenshots private and avoid sending raw images to cloud models, enable the local Ollama vision mode:

{
  "plugin": [
    ["opencode-desktop-context", {
      "visionModel": "moondream:latest",
      "ollamaBaseUrl": "http://127.0.0.1:11434",
      "autoAttach": true,
      "periodicCaptureMs": 300000
    }]
  ]
}

When visionModel is set:

• autoAttach sends a text description of the desktop instead of the image file.
• The plugin exposes a describe_desktop tool the assistant can call explicitly.
• Screenshots are still captured periodically to keep context fresh.

Install a vision model in Ollama first:

ollama pull moondream:latest

Development

bun install
bun test
bun run build

License

MIT — see LICENSE.

Disclaimer

This project is an independent community plugin. It is not affiliated with, endorsed by, or maintained by OpenCode or its makers. All product names, trademarks, and registered trademarks mentioned in this documentation are the property of their respective owners.