Directory entry
liustack
modlens plugin: Vision for DeepSeek Harness Text-Only Models
A vision bridge for DeepSeek Harness and text-only coding agents that returns structured OCR, layout, and semantic evidence from images.
What this plugin does
The modlens plugin is a vision bridge for DeepSeek Harness, enabling text-only models to read images pasted directly into chat. It converts images into structured JSON evidence, including OCR, layout regions, and semantic details, without requiring file paths.
The modlens plugin is the first vision plugin for DeepSeek Harness, designed to give text-only models sight. It reads images pasted directly into the chat, converting them into structured JSON evidence that includes full transcription, reading-order layout regions, and entity and relation lists. The plugin reuses existing multimodal model setups from tools like Claude Code, Codex, OpenCode, or Pi, and can also use free options like Antigravity CLI or a Gemini key. It is verified on real machines in several harnesses, and the modlens plugin adds vision entries to the model selector for DeepSeek routes.
Documented capabilities
Image-to-JSON Evidence
The modlens plugin converts pasted images into structured JSON evidence, including OCR transcription, layout regions, and semantic details, enabling text-only models to ground answers in visual content.
Zero-Config Integration
The modlens plugin reuses existing multimodal model configurations from Claude Code, Codex, OpenCode, or Pi, requiring no additional setup if those tools are already configured.
Multi-Harness Support
The modlens plugin is verified on real machines in Claude Code, Codex, Pi, and OpenCode, allowing the same vision capability to be used across different harness environments.
Practical use cases
Reading Screenshots
Users can paste screenshots into DeepSeek Harness, and the modlens transcribes the content, including UI elements, text, and layout, for text-only models to analyze.
Analyzing Charts
The modlens can read dense charts, such as scatter plots, extracting axis labels, scales, and highlighted regions, which are often challenging for vision bridges.
Processing Multiple Images
Users can paste multiple images at once, and the modlens reads them sequentially, identifying visual families and describing each image's content and style.
Review before you install
Read the source, recent releases, requested permissions, network access, and maintainer history before trusting a third-party plugin.
Frequently asked questions
How does the modlens handle pasted images?
The modlens reads images pasted directly into the chat, converting them to structured JSON evidence at request time. The original thumbnail is preserved, and the model answers based on the extracted evidence.
What vision engines does the modlens support?
The modlens reuses multimodal models from Claude Code, Codex, OpenCode, or Pi if available. It also supports free options like Antigravity CLI and Gemini API keys for faster reads.
Is the modlens compatible with all DeepSeek models?
The modlens wraps DeepSeek and GLM text-only models. DeepSeek's own vision models are automatically excluded from the wrap, as they already have vision capability.