codex-vision-proxy Takes 181 Stars in a Day Giving Text-Only Models Working Access to Codex's view_image Tool
Anionex/codex-vision-proxy launched 2026-08-01 and hit 181 stars in 24 hours by solving a mundane compatibility gap: text-only models such as DeepSeek cannot consume images when Codex calls its built-in view_image. The fix is a local proxy on 127.0.0.1:19100 that intercepts image-bearing requests, converts them to text descriptions via a separate vision API, and forwards the request with the original auth header untouched, so only vision credentials are added. It processes multiple images per request in parallel, caches descriptions by SHA256, and routes through scene-specific templates using preceding user text as a focus hint, so error screenshots yield stack traces and UI screenshots yield element lists with positions. Two CLIs ship alongside: glance for image Q&A and OCR, ground for returning pixel coordinates from a natural-language query.
Source
↳ Follow the thread