Codex Deepseek Vision

codex-deepseek-vision

[](LICENSE) [](https://www.python.org/) [](https://github.com/SIMON-WORLD/codex-deepseek-vision/releases)

**Codex vision bridge for DeepSeek V4 Flash and other text-only models.** Package/CLI: `agent-vision`. DeepSeek V4 Flash now speaks the Responses protocol and runs inside Codex/ChatGPT, but the model itself is text-only and cannot see images. agent-vision is a free local vision proxy: pasted images and `view_image` calls are converted into text through an OpenAI-compatible vision API (free GLM-4V-Flash by default) before DeepSeek reasons. No Ollama, no GPU, no model swap.

**English** | [中文](README.zh-CN.md)

Why

Text-only agents cannot see pasted screenshots, local images, charts, or error dialogs. Replacing the model usually means paying more or changing your whole workflow. agent-vision sits between the agent and its model provider and does the conversion automatically:

• Paste an image in your agent, and the local proxy rewrites it into text before the request reaches the text-only model.
• Ask the agent to inspect a local image, an image URL, or the latest image you pasted (see --latest), and it returns a factual description.
• Keep your existing model, key, and workflow. Everything is local, reversible, and free by default.

Architecture

flowchart LR
  U[User pastes an image] --> A[Any AI agent]
  A -->|request with image| P[agent-vision proxy :19100]
  P --> V[OpenAI-compatible vision API]
  V -->|text description| P
  P -->|text-only request| M[DeepSeek / text-only model]

`see` mode skips the proxy: the image path is sent directly to the vision API and the returned text is used by the agent.

Supported Agents

Agent	Integration	Status
Codex	Safe auto-patch: rewrit