Opencode Vision

Opencode Vision is a plugin for OpenCode that enables vision support for models that lack native image attachment support.

Originally built for MiniMax models, it can be configured to work with any model that requires MCP-based image handling.

It restores the "paste and ask" workflow by automatically saving image assets and routing them through a configured MCP image analysis tool such as local_vision.

Compared with @devadathanmb/opencode-minimax-easy-vision, it doesn't save images from the clipboard to the local device.

Table of Contents

• Demo
• The Problem
• Prerequisites
• Installation
• What This Plugin Does
• Supported Models
  • Custom Model Configuration
    • Locations (Priority Order)
    • Config Format
    • Pattern Syntax
    • Wildcard Rules
    • Configuration Examples
  • Custom Image Analysis Tool
  • Custom Prompt Template
• Supported Image Formats
• Usage
  • Example Interaction
• Development
• License
• References

Demo

See how it works:

https://github.com/user-attachments/assets/df396c6c-6fa8-46b8-8984-c003ecf1c12c

https://github.com/user-attachments/assets/826f90ea-913f-427e-ace8-0b711302c497

The Problem

When using MiniMax models (like MiniMax M2.1) in OpenCode, native image attachments aren't supported.

These models usually need an MCP image analysis tool with an explicit file path. This breaks the normal flow:

• Ignored images: Pasted images are simply ignored by the model.
• Manual steps: You have to save screenshots manually, find the path, and reference it in your prompt.
• Broken flow: The "paste and ask" experience available with Claude or GPT models is lost.

Prerequisites

The MCP server that exposes the local_vision tool must be configured in your opencode.json before using this plugin.

Installation

Add the plugin to the plugin array in your opencode.json file:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-vision"]
}

What This Plugin Does

This plugin automates the vision pipeline so you don't have to think about it.

How it works:

1. Detects when a configured model is active.
2. Intercepts images pasted into the chat.
3. Saves them to a temporary local directory.
4. Injects the necessary context for the model to invoke the configured image analysis tool with the correct path.

Result: You just paste the image and ask your question just like how you do with Claude or GPT models. The plugin handles the rest.

Supported Models

By default, the plugin activates for models matching the pattern ["*/qwen3-coder-next-mlx"]:

• */qwen3-coder-next-mlx — the qwen3-coder-next-mlx model from any provider

Custom Model Configuration

You can enable this for other models by creating a config file.

Locations (Priority Order)

1. Project level: .opencode/opencode-vision.json
2. User level: ~/.config/opencode/opencode-vision.json

Config Format

{
  "models": ["*/qwen3-coder-next-mlx", "z-ai/*", "*/minimax-m2.5"]
}

Pattern Syntax

Wildcard Rules

• *suffix matches values ending with suffix
• prefix* matches values starting with prefix
• * matches everything
• *text* matches values containing text

If the config is missing or empty, it defaults to ["*/qwen3-coder-next-mlx"].

Configuration Examples

Enable for all models:

{
  "models": ["*"]
}

Specific providers:

{
  "models": ["minimax/*", "z-ai/*", "google/*"]
}

Mix of providers and models:

{
  "models": ["minimax/*", "z-ai/glm-4.7", "*/minimax-m2.5"]
}

Custom Image Analysis Tool

By default, the plugin uses the local_vision tool for image analysis. You can still configure a different tool if needed.

[!NOTE] The imageAnalysisTool value is the tool name, not the MCP server name.

MCP servers expose one or more tools. Set imageAnalysisTool to the exact tool name you want the model to call.

Example: openrouter-image-mcp

openrouter-image-mcp routes image analysis through OpenRouter, giving you access to any vision-capable model including free ones.

Add the MCP server to your opencode.json:

{
  "mcp": {
    "openrouter_image": {
      "type": "local",
      "command": ["npx", "openrouter-image-mcp"],
      "environment": {
        "OPENROUTER_API_KEY": "your-api-key-here",
        "OPENROUTER_MODEL": "nvidia/nemotron-nano-12b-v2-vl:free"
      }
    }
  }
}

Then configure the plugin to use it:

{
  "models": ["*/qwen3-coder-next-mlx"],
  "imageAnalysisTool": "mcp_openrouter_image_analyze_image"
}

[!TIP] nvidia/nemotron-nano-12b-v2-vl:free is a free vision model on OpenRouter that requires no credits.

Custom Prompt Template

By default, the plugin generates a fixed instruction prompt telling the model to use the image analysis tool. You can override this with a custom template:

{
  "promptTemplate": "I'm attaching {imageCount} image(s) for you to analyze.\n\nImages:\n{imageList}\n\nUse the `{toolName}` tool on each one.\n\nMy question: {userText}"
}

Available Variables

The template must contain at least one variable — if none are present, the plugin falls back to the default prompt.

Supported Image Formats

• PNG
• JPEG
• WebP

(The plugin currently accepts PNG, JPEG, and WebP inputs and passes them to your configured image tool.)

Usage

1. Select a supported model in OpenCode.
2. Paste an image (Cmd+V / Ctrl+V).
3. Ask a question about it, just like how you do for other models with native vision support.

Example Interaction

You: [pasted screenshot] Why is this failing?

Model: I'll check the image using the local_vision tool. [Calls local_vision path="/tmp/xyz.png"]

Model: The error suggests a syntax error on line 12.

Development

1. Install dependencies and build:

   npm install
   npm run build
   

2. The built plugin will be available at dist/index.js.
3. Symlink it into the global plugin directory and restart OpenCode to pick up changes:

   mkdir -p ~/.config/opencode/plugin
   ln -sf $(pwd)/dist/index.js ~/.config/opencode/plugin/opencode-vision.js
   

License

AGPL-3.0. See LICENSE

References

• OpenCode Official Website
• OpenCode Plugins Documentation
• MiniMax Official Website
• MiniMax Coding Plan MCP Repository
• MiniMax API Documentation
• opencode-minimax-easy-vision