You paste an image into your OpenCode session. The LLM can't see it. The plugin can.

opencode-vision-paste intercepts pasted images, sends them to a local VL API (e.g., Qwen3VL running via llama.cpp), and replaces the image with a detailed text description — seamlessly, before the LLM ever sees the message.

No cloud dependencies. No API keys required. Your images stay local.

---

Features

• Transparent — images are silently converted to text. The LLM never knows there was an image
• Zero dependencies — single ESM file, no build step, no npm deps
• Multi-image support — paste multiple images at once; each is analyzed and numbered
• Local-first — works with any OpenAI-compatible VL API (llama.cpp, vLLM, etc.)
• Smart caching — reuses analysis results for duplicate images across conversation turns
• HTTP & data URLs — handles both pasted images (data: URLs) and web-hosted images
• Smart skip — automatically detects when the current chat model natively supports images and skips interception

Installation

For Humans

Copy and paste this prompt to your OpenCode agent:

Install and configure opencode-vision-paste by following the instructions here:
https://raw.githubusercontent.com/wsaaaqqq/opencode-vision-paste/main/readme/INSTALL.md

For LLM Agents

curl -s https://raw.githubusercontent.com/wsaaaqqq/opencode-vision-paste/main/readme/INSTALL.md

Or read the Installation Guide directly.

---

Quick Start

1. Install the plugin (see Installation above)
2. Configure your VL API endpoint in .opencode/vision-paste.config.jsonc (optional — defaults work if you have a local llama.cpp running)
3. Paste an image in any OpenCode chat — watch it turn into text

---

Configuration

All settings are optional. The plugin works with sensible defaults.

Option	Default	Description
apiBaseUrl	http://192.168.9.44:5678/v1	OpenAI-compatible VL API endpoint
apiModel	Qwen3VL-8B-Instruct-Q4_K_M.gguf	Model name for the VL API
apiKey	""	API key (leave empty if not required)
promptTemplate	""	Custom prompt (overrides promptLocale); {userText} is replaced with the user's message
promptLocale	zh	Built-in prompt language: en, zh, ja, ko, es, fr, de, ru, pt
skipIfModelSupportsVision	true	Skip interception when the current chat model natively supports images
visionModels	[]	Extra model ID patterns to treat as vision-capable (case-insensitive)
healthCheckOnStart	true	Ping VL API on session start; warn if unreachable
verbose	false	Show analysis progress in chat messages
errorHints	true	Include troubleshooting suggestions in error messages

Config file locations (first found wins):

1. .opencode/vision-paste.config.jsonc (project-level)
2. ~/.config/opencode/vision-paste.config.jsonc (user-level)

Full reference: CONFIGURATION.md

---

How it works

User pastes image
       ↓
opencode-vision-paste intercepts `experimental.chat.messages.transform`
       ↓
Decodes image (data URL or HTTP) → saves temp file
       ↓
Sends to local VL API (OpenAI-compatible chat completions)
       ↓
Replaces image part with analysis text
       ↓
Temp file deleted — LLM sees text only

The plugin hooks into OpenCode's experimental.chat.messages.transform pipeline, running before the message is sent to the LLM.

---

VL API Setup

The plugin needs a running VL API. Choose one:

Docker (one command)

# 1. Download the model
mkdir models
wget -P models https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct-GGUF/resolve/main/Qwen3VL-8B-Instruct-Q4_K_M.gguf

# 2. Start the server
docker compose up -d

Ollama

ollama pull qwen2.5-vl:7b
# Then configure apiBaseUrl: http://localhost:11434/v1

See examples/ for configuration presets.

---

CLI

npx opencode-vision-paste init     # Interactive setup wizard
npx opencode-vision-paste doctor   # Diagnose plugin + VL API
npx opencode-vision-paste config   # Show current configuration

---

FAQ

Q: Images are not being analyzed? Run npx opencode-vision-paste doctor to check your setup. Common issues: VL API not running, wrong URL, or model not loaded.

Q: How do I use a different VL API backend? Update apiBaseUrl in .opencode/vision-paste.config.jsonc. Any OpenAI-compatible endpoint works (Ollama, vLLM, llama.cpp, LM Studio, etc.).

Q: Can I force the plugin to always process images? Set "skipIfModelSupportsVision": false in your config.

Q: The prompt is in Chinese, can I change it? Set "promptLocale": "en" or provide a custom "promptTemplate".

---

Contributing

PRs welcome! See CONTRIBUTING.md.

---

License

MIT