ModLens

Overview

ModLens solves a specific gap in the agent ecosystem: many powerful coding models are text-only, but developers constantly need to share visual context — UI screenshots, error dialogs, charts, architecture diagrams, whiteboard photos. ModLens bridges this by converting pasted images into structured JSON evidence that any text-only model can process. The agent receives OCR transcriptions, layout regions in reading order, and entity/relation lists — enough structured data to reason accurately about visual content without native vision capabilities.

Key Capabilities

Zero-friction image input: Users paste images directly into chat. ModLens auto-detects when a text-only model is active and intercepts the image before the model sees it, converting it to structured evidence. Native vision models are excluded automatically — no unnecessary processing when the model can already see.

Failover architecture: ModLens routes image processing through a chain of up to ten vision providers. Built-in providers include Gemini API, OpenAI-compatible endpoints (Qwen-VL, GLM, self-hosted gateways), Anthropic API, and Antigravity CLI. It also reuses existing CLI logins from Claude Code, Codex, Kimi Code, and Grok — no additional API keys needed if you already have one of those tools authenticated. If the first provider fails, the request automatically falls through to the next.

Structured evidence, not summaries: Instead of returning a prose description of an image (“this appears to be a form with three fields”), ModLens returns structured JSON with full transcription text, layout regions tagged with reading order and spatial coordinates, and entity/relation lists. This gives the text-only model precise, quotable evidence to work with rather than generated assumptions.

Dense content handling: ModLens successfully processes complex visuals that trip up simpler approaches — scatter plots with overlapping labels, multi-element UIs with nested components, dense documentation pages, and social media posts with embedded media.

Use Cases

Developers working with DeepSeek or other text-only models install ModLens to share UI screenshots during debugging sessions — the agent can read error messages, identify layout issues, and reference specific elements by position. Teams use it to extract data from charts and visualizations that need to be discussed in code context. It handles document OCR for specification pages, whiteboard photos for architecture discussions, and social media content analysis — any scenario where a developer would normally say “look at this screenshot” but the model cannot see.

Considerations

Image processing adds 5-45 seconds of latency depending on the provider used (built-in API providers are faster at 5-10 seconds; reused CLI logins take 15-45 seconds). The quality of extracted evidence depends on the vision provider in the failover chain — Gemini API and Anthropic tend to produce the most accurate structured output. ModLens is MIT-licensed and the author accepts issue reports but declines pull requests.

Who It’s For

ModLens is for developers using text-only coding agents (particularly DeepSeek-based) who regularly need to share visual context. It is especially valuable for teams that chose a text-only model for cost or performance reasons but do not want to give up the ability to discuss screenshots, charts, and UI mockups with their agent.