These days I reach for Codex CLI more than Claude Code. Not just for token economics, Codex has been catching up fast and is already ahead in some areas. It was open source from day one, arguably better code quality than the leaked Claude Code source, yet nobody cared. It's been improving quietly under the radar.
I subscribe to ChatGPT Pro, but Ralph + Fast Mode can still burn through tokens. My company has a sizable Cloudflare credit from the Startups program (https://www.cloudflare.com/forstartups/) that we're barely using. Workers AI has some interesting models like Kimi K2.5, Gemma 4, GLM-4.7-Flash, GPT-OSS-120B, so I wanted to route Codex through them.
Tried OpenCode's Cloudflare provider first but Codex's UX fits me better and the integration felt incomplete. Workers AI has an OpenAI-compatible API so I expected it to just work with Codex. Nope. The Responses API surface doesn't map cleanly. A local proxy was the simplest fix so I had Codex write one.
It's a small TypeScript proxy that translates between Codex CLI and Workers AI. Clone, add credentials, npm start, done. There's a shell wrapper (codexcf) that handles daemon management and model catalog automatically.
Been running Kimi K2.5 for a day, works well enough for my workflow. Gemma 4 looks promising too, and Cloudflare told me Qwen 3.5 is coming soon. GLM 5.1 probably not far behind.
If you have Cloudflare credits collecting dust, this might be a good way to spend them.
These days I reach for Codex CLI more than Claude Code. Not just for token economics, Codex has been catching up fast and is already ahead in some areas. It was open source from day one, arguably better code quality than the leaked Claude Code source, yet nobody cared. It's been improving quietly under the radar.
I subscribe to ChatGPT Pro, but Ralph + Fast Mode can still burn through tokens. My company has a sizable Cloudflare credit from the Startups program (https://www.cloudflare.com/forstartups/) that we're barely using. Workers AI has some interesting models like Kimi K2.5, Gemma 4, GLM-4.7-Flash, GPT-OSS-120B, so I wanted to route Codex through them.
Tried OpenCode's Cloudflare provider first but Codex's UX fits me better and the integration felt incomplete. Workers AI has an OpenAI-compatible API so I expected it to just work with Codex. Nope. The Responses API surface doesn't map cleanly. A local proxy was the simplest fix so I had Codex write one.
It's a small TypeScript proxy that translates between Codex CLI and Workers AI. Clone, add credentials, npm start, done. There's a shell wrapper (codexcf) that handles daemon management and model catalog automatically.
Been running Kimi K2.5 for a day, works well enough for my workflow. Gemma 4 looks promising too, and Cloudflare told me Qwen 3.5 is coming soon. GLM 5.1 probably not far behind.
If you have Cloudflare credits collecting dust, this might be a good way to spend them.