Repolode
Submit
liustack

liustack/modlens

A vision plugin for DeepSeek Harness that lets text-only models like DeepSeek and GLM read pasted images as structured JSON (OCR, layout, semantics).

View on GitHubWebsite
Stars
4,398
Forks
133
Open issues
6
Last commit
7 days ago

What is it?

ModLens gives text-only language models the ability to read images. Many DeepSeek and GLM chat models cannot accept image input directly, so ModLens sits between the chat and a vision engine: when an image is pasted, it is sent to a vision provider and the result comes back as structured evidence covering full transcription, reading-order layout regions, and entity and relation lists.

It works as a plugin for DeepSeek Harness (dsh) and as an installable agent skill for Claude Code, Codex, OpenCode, and Pi. Pasted images are handled through a private temp file path or through a model-selector entry that wraps the same route. Vision engines include a free Gemini API key, any OpenAI-compatible endpoint, an Anthropic key, and local agent CLIs whose existing logins can be reused with explicit per-harness consent. Multiple keys can be comma-separated and rotate on auth, rate-limit, or quota errors.

The project is MIT-licensed TypeScript, installed through npm as @liustack/modlens and configured via a CLI with a doctor health check.

Key features

  • 01Reads images pasted into the chat and returns structured JSON with OCR, layout, and semantics.
  • 02Works with text-only DeepSeek and GLM models in DeepSeek Harness, Claude Code, Codex, OpenCode, and Pi.
  • 03Supports a free Gemini key, any OpenAI-compatible vision endpoint, Anthropic, and reusable local agent CLIs.
  • 04Fails over across configured engines and records every attempt in meta.attempts.
  • 05Rotates comma-separated API keys on auth, rate-limit, or quota failures.
  • 06Includes a CLI with config commands and a doctor health check.

Use cases

  • Letting a DeepSeek or GLM chat session discuss screenshots, slides, or tweets pasted into it.
  • Extracting text and layout from dense charts or documents inside a coding agent workflow.
  • Driving a vision model that exposes an OpenAI-compatible API as the image reader for a text-only agent.
  • Reusing an existing Claude Code, Codex, or Pi login for image reading without adding new API keys.

Topics

#vision#ocr#multimodal#deepseek#agent-skills#image-to-text#llm-tools#claude-code#claude-skills#codex

FAQ

What is modlens?
A vision plugin for DeepSeek Harness that lets text-only models like DeepSeek and GLM read pasted images as structured JSON (OCR, layout, semantics). ModLens gives text-only language models the ability to read images. Many DeepSeek and GLM chat models cannot accept image input directly, so ModLens sits between the chat and a vision engine: when an image is pasted, it is sent to a vision provider and the result comes back as structured evidence covering full transcription, reading-order layout regions, and entity and relation lists.
Is modlens free to use? What license does it have?
modlens is open source under the MIT license. Check the license terms before using it in commercial products.
Is modlens still maintained?
Actively maintained. The last commit to modlens was on Oct 15, 2026.
What can I use modlens for?
Letting a DeepSeek or GLM chat session discuss screenshots, slides, or tweets pasted into it. Extracting text and layout from dense charts or documents inside a coding agent workflow. Driving a vision model that exposes an OpenAI-compatible API as the image reader for a text-only agent. Reusing an existing Claude Code, Codex, or Pi login for image reading without adding new API keys.
Where can I get modlens?
The source code and installation instructions are on GitHub: https://github.com/liustack/modlens

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects