codex-imagegen

Route OpenClaw (or any agentic tool that can shell out) image generation through your ChatGPT Pro/Plus subscription instead of the OpenAI API.

Wraps Codex CLI's built-in $imagegen skill so gpt-image-2 runs bill against your Codex subscription credits — not your OpenAI API key. If you already pay for ChatGPT Pro ($200/mo) or Plus ($20/mo), this turns that subscription into the image generator your agents already wanted.

---

Why this exists

OpenClaw's native image_generate tool resolves to api.openai.com and requires a paid API key. If you already have a ChatGPT subscription you're effectively paying twice — once for the plan, again per image. Codex CLI has $imagegen baked in but it's behind a feature flag marked "under development," and no third-party framework surfaces it by default. This skill closes that gap.

Requirements

• Codex CLI installed and signed in with your ChatGPT Pro/Plus account (codex login)
• bash available (Git Bash on Windows, native on macOS/Linux)
• Any agentic framework that can invoke shell scripts (OpenClaw, Claude Code, custom agents, etc.)

Install

1. Copy the codex-imagegen/ folder into your workspace's skills directory. For OpenClaw: <workspace>/skills/codex-imagegen/.
2. Make the bash script executable (macOS/Linux):

   chmod +x codex-imagegen/scripts/imagegen.sh
   

3. Enable the Codex image feature flag (one-time — persists in ~/.codex/config.toml):

   codex features enable image_generation
   

4. Point your agent at it. For OpenClaw, add to TOOLS.md:

   Image generation: always use bash skills/codex-imagegen/scripts/imagegen.sh. Never call the built-in image_generate tool — that hits the paid API path.

5. (Recommended) Add a matching hard rule to AGENTS.md so every new session boot reinforces it.

Usage

bash codex-imagegen/scripts/imagegen.sh \
  --prompt "cinematic cat astronaut on a red dune at sunset" \
  --size 1024x1024 \
  --quality medium \
  --out /absolute/path/output.png

• --size: 1024x1024 | 1024x1536 (portrait) | 1536x1024 (landscape) | 2000x2000
• --quality: low | medium (default) | high — high burns credits ~3–5× faster
• --ref <path>: repeatable; passes the image to codex exec -i as a style/identity anchor
• Stdout on success = the final PNG path
• Default output dir = <skill-dir>/output/; override with $CODEX_IMAGEGEN_OUT_DIR or --out

Caveats (honest)

• The image_generation feature flag is still marked "under development" in Codex CLI. It works today; could change in a future release.
• Subscription-billed image generation burns Codex credits 3–5× faster than a text turn. ChatGPT Pro has a 10× Codex multiplier through May 31, 2026 — substantial headroom but not unlimited.
• No video support. Codex CLI has no $videogen. Use sora.com manually (Pro = unlimited) or the paid sora-2 API.
• OpenAI could remove subscription access for this path at any time. Bet on it accordingly.

Troubleshooting

• "image_generation feature disabled": codex features enable image_generation, retry.
• "no PNG found in thread dir": the inner Codex turn didn't invoke $imagegen. Prefix your prompt with Use the $imagegen skill to generate: to force it.
• Windows sandbox -1073741502 errors: cosmetic noise from the inner Codex session's shell attempts. Doesn't block image generation; the wrapper ignores them.
• Want to fall back to the API path? Set OPENAI_API_KEY and your framework's built-in image tool will still work. They're independent paths.

License

MIT. See LICENSE.

Contributing

Issues and PRs welcome. Especially:

• macOS / Linux install verifications (Windows is the primary tested path)
• Support for other agent frameworks beyond OpenClaw (wrappers, adapters)
• Reference-image fidelity improvements if you figure out better $imagegen prompting

---

Built while trying to stop hemorrhaging API dollars on an automated NFT content pipeline. If you're doing something similar and find another gap like this, tell me.