liustack/modlens

What it solves

ModLens allows text-only AI models (such as DeepSeek and GLM) to "see" and process images. It eliminates the need for users to manually save images to files and provide paths, allowing them to simply paste images directly into the chat interface.

How it works

It acts as a vision plugin or "bridge" that intercepts pasted images and sends them to a separate vision engine for processing. The engine converts the image into structured evidence—including full transcriptions, layout regions, and entity lists—which is then fed back to the text-only model. It supports a wide range of vision providers, including Gemini, OpenAI-compatible endpoints, Anthropic, and existing local agent CLIs (like Claude Code or Codex), and can automatically failover between them to ensure a result is returned.

Who it’s for

Developers and users of text-only LLM harnesses (such as DeepSeek Harness, Claude Code, Codex, Pi, and OpenCode) who want multimodal capabilities without switching to a different model or manually managing image files.

Highlights

• Native Integration: The first vision plugin for DeepSeek Harness (dsh), allowing images to be read via a modlens_read_image tool.
• Zero-Config Setup: Can reuse existing logins from other AI CLIs on the machine or use free options like the Gemini API or Antigravity CLI.
• Structured Output: Provides evidence-based descriptions (transcriptions and layout) rather than imaginative summaries.
• Flexible Engine Pool: Supports six built-in providers and four reusable local CLIs with an automatic failover chain.
• Lightweight: Installs as a single folder or plugin with no requirement for local proxy daemons or harness configuration changes.