opencode-glm-clipboard

OpenCode plugin that fixes pasted clipboard image flow for text-only GLM models (GLM-5.x, GLM-4.6, etc.).

What it does

When a clipboard image arrives on a model that can't read images natively, this plugin:

1. Saves the image to a temporary local file (instant, non-blocking)
2. Replaces the image part with a text marker: [Image saved to: /path/to/file.png]
3. Injects a system-prompt instruction telling the model to read the file on-demand via the configured zai-mcp-server MCP tools (extract_text_from_screenshot for text/code/errors, analyze_image for UI/diagrams/photos)

The model calls the vision MCP tools itself when it needs to understand the image — the plugin never blocks the chat to run OCR.

Handles pasted images delivered either as embedded data URLs (Cmd+V) or as file-path attachments (Ctrl+V / dragged files).

Why

GLM models served via OpenCode reject direct image input. OpenCode's built-in read tool also hard-refuses images for non-vision models. This plugin bridges the gap: it saves the image and instructs the model to use the already-running zai-mcp-server MCP to read it on-demand.

Prerequisite: zai-mcp-server MCP

The system nudge references zai-mcp-server_extract_text_from_screenshot and zai-mcp-server_analyze_image. You need a local stdio MCP server named zai-mcp-server in your OpenCode config:

{
  "mcp": {
    "zai-mcp-server": {
      "type": "local",
      "command": ["npx", "-y", "@z_ai/mcp-server"],
      "environment": {
        "Z_AI_API_KEY": "<your-key>",
        "Z_AI_MODE": "ZAI"
      }
    }
  }
}

If the MCP server is absent, the model will still see the [Image saved to: ...] marker but won't be able to read the image (the nudge will reference tools that don't exist).

Install

npx -y opencode-glm-clipboard@latest

Then restart OpenCode.

Uninstall

npx -y opencode-glm-clipboard@latest --uninstall

Local development

npm install
npm run build
npm test

Configuration

• OPENCODE_GLM_CLIPBOARD_MAX_AGE_HOURS
  • Optional. Default: 24.
  • Controls temp image cleanup threshold.

Temp directory used: ~/.cache/opencode/pasted-images/ (or $OPENCODE_GLM_CLIPBOARD_TEST_DIR in tests).

Compatibility

• Acts on models from providers that don't declare native image input support:
  • CrofAI provider (e.g. deepseek-v4-pro)
  • zai-coding-plan provider (e.g. glm-5.2)
  • Legacy glm-* model IDs
• Leaves image-capable models untouched (e.g. CrofAI/kimi-k2.6, CrofAI/kimi-k2.6-precision)
• Leaves other providers/models untouched

Release flow

Recommended release order:

1. Bump version in package.json
2. Commit and push main
3. Create and push tag vX.Y.Z
4. Publish GitHub Release for that tag
5. GitHub Actions publishes to npm automatically

This repository includes a workflow that publishes to npm when a GitHub Release is published.

Recommended setup: npm Trusted Publishing (OIDC), no long-lived NPM_TOKEN required.

Add local plugin path in your OpenCode config if testing without publish:

{
  "plugin": [
    "file:///absolute/path/to/opencode-glm-clipboard/dist/index.js"
  ]
}

License

MIT