dsh-codex-media

DeepSeek multimodal models can already read images, yet nothing in the Harness conversation loop could draw them. dsh-codex-media bridges that gap by routing image generation and editing through your local Codex CLI, which draws from a ChatGPT Plus subscription. Say paint a cyberpunk cat and the model calls generate_image, writing the file to your working directory; paste a reference and ask to change the palette and edit_image redoes the render with that reference attached. No browser tab, no manual API calls.

Under the hood the plugin registers a stdio MCP server exposing four tools: generate_image renders to a path you specify, edit_image accepts a local file path or clipboard reference for guided redraws, clipboard_to_image saves a screenshot to disk via pwsh -STA on Windows, and codex_task delegates a self-contained scripting or file-manipulation job to Codex so it does not consume DeepSeek tokens. Image understanding stays on the native vision path; the plugin does not duplicate that work.

It is built for developers who already run a ChatGPT Plus subscription and have Codex CLI installed, and want in-conversation image output from Harness. If your day includes quick concept renders, tweaking UI screenshots, batch-saving clipboard captures, or offloading a small script task to protect the main model budget, a single natural-language prompt covers it all. Non-Windows platforms get the full generation and editing experience; only the clipboard input tool is unavailable there.