A vision sidecar for text-only LLMs

What it does claude-vision-skill is a thin JavaScript shim that gives non-vision models—such as DeepSeek—a way to process images. When a user drops a picture into the chat, vision.js base64-encodes it and fires it off to a vision-capable API like Alibaba’s Qwen or OpenAI’s GPT-4o-mini. The resulting text description is fed back to the main model as context, letting it respond as if it understood the image itself.

The interesting bit The repository is as much a prompt-engineering hack as it is code. It ships a CLAUDE.md instruction file that teaches Claude Code when to invoke the script, effectively turning the AI assistant into its own installer. The README even splits its instructions into sections explicitly addressed to AI readers and humans—a rare admission that the documentation itself will be parsed by both.

Key highlights

• Dead-simple architecture: one vision.js script plus a prompt file.
• Vendor-agnostic via OpenAI-compatible API format; swap BASE_URL and model name to change providers.
• Cost-optimized for Chinese users: recommends Alibaba Cloud Qwen models with a 1-million-token free tier over pricier OpenAI options.
• Includes a separate integration guide for cyberboss, a WeChat bot framework.
• Roughly 1,800 stars suggest plenty of developers are running vision-less models and want a quick fix.

Caveats

• The entire README and setup flow is in Chinese, so non-Chinese speakers will need translation.
• It is fundamentally glue code: a base64 wrapper around an API call, not a model or training pipeline.
• Requires bringing your own API key and manually editing the script for credentials.

Verdict Grab this if you are self-hosting or using a text-only model via Claude Code and need image understanding without switching models. If you are already on GPT-4o, Claude 3.5 Sonnet, or another native vision model, there is nothing here to see.