跪拜 Guibai
← All articles
AI Coding · OpenAI

A New OpenCode Plugin Gives Text-Only LLMs Vision by Routing Images to a Multimodal Model

By 晴天小庭 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Many developers prefer open-source coding models for cost or privacy, but the strongest ones remain text-only. This plugin removes the vision ceiling without replacing the primary model, making it practical to stick with DeepSeek or GLM for coding workflows that occasionally need image understanding.

Summary

Open-source coding assistants often rely on strong but text-only models like DeepSeek and GLM, which break down the moment a task requires reading a screenshot or diagram. The analyze-image plugin sidesteps that limitation by injecting a tool into the chat context. When the primary model detects an image it cannot process, the tool forwards it to a separate, configurable vision model and feeds the resulting text description back into the conversation.

Users maintain a configurable list of models that need the image-routing treatment, so the tool only activates for text-only LLMs. The plugin is open source but has not yet passed OpenCode's official ecosystem verification, so installation currently requires manually placing the source code into a designated OpenCode directory.

CipherTing, the plugin's creator, points users to GitHub issues for bug reports and the Juejin comments section for broader discussion.

Takeaways
Text-only models like DeepSeek and GLM cannot process images, blocking any coding task that involves screenshots or diagrams.
The analyze-image plugin adds a tool to the OpenCode chat context that detects image prompts and routes them to a separate vision model.
Users configure which models should trigger the image-routing behavior, so vision-capable models are left untouched.
The vision model's text output is inserted into the conversation, letting the primary text-only LLM continue as if it had seen the image.
Installation currently requires manually copying source code into a specific OpenCode folder because the plugin has not passed official ecosystem verification.
Conclusions

OpenCode's plugin architecture is flexible enough to retrofit multimodal capabilities onto text-only models through tool-calling, which is a lighter lift than waiting for every open-source model to ship native vision support.

The plugin's configurable model list is a pragmatic design choice: it avoids wasting tokens and latency by routing images through a vision model when the primary LLM can already handle them natively.

Concepts & terms
OpenCode
An open-source AI coding assistant that supports plugins and multiple language models.
Multimodal plugin
A plugin that enables a text-only language model to handle images by routing them to a separate vision-capable model and returning a text description.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗