Claude Code can't generate images, and everything it does runs on your Claude quota. The Codex CLI can generate images with gpt-image-2, and runs on your ChatGPT plan — a completely separate budget. This plugin bridges the two.

Two things you get:

1. Claude can produce images. Ask for an icon, a mockup, a hero image, a full favicon set — Claude writes the prompt, drives gpt-image-2 through Codex, then looks at the result and regenerates if it missed.

2. Claude can delegate to GPT-5. Four Codex-powered subagents handle review, debugging, bulk implementation, and image work. Claude stays the orchestrator: it scopes the task, Codex does the volume, Claude verifies the result. Heavy work runs on your ChatGPT quota and its intermediate output never enters your Claude context.

No OpenAI API key. Everything routes through codex login, so it bills against your ChatGPT plan rather than API credits.

Demo

One style contract across a set. Each sibling was generated with an earlier icon passed as --ref, so the palette, stroke weight, glyph scale and margins carry over instead of being re-described and re-interpreted every time:

Every image in this repository was generated by this plugin — the GIF above, these icons, and the social preview card. The exact commands and the VHS recording script are in assets/README.md, so you can reproduce all of it.

Contents

• Demo · Install · Quickstart · What's inside
• Feature 1 — images · Feature 2 — GPT-5 subagents
• How it works · Limitations · Troubleshooting

Install

Prerequisites

Then:

/plugin marketplace add Sateezg/codex-bridge
/plugin install codex-bridge@codex-bridge

Run /reload-plugins if Claude Code asks you to.

git clone https://github.com/Sateezg/codex-bridge.git
claude --plugin-dir ./codex-bridge

git clone https://github.com/Sateezg/codex-bridge.git ~/.claude/skills/codex-bridge
chmod +x ~/.claude/skills/codex-bridge/bin/*

Quickstart

Nothing to configure. Just ask in plain language:

generate a hero image for the landing page, save it in assets/
make the sky in assets/hero.png sunset orange, keep everything else
set up favicons and an OG card for this project
review my changes before I commit
ask codex why this test only fails under --parallel
rename useSession to useAuthSession across the whole repo

The last one is the interesting case. Claude notices it's mechanical work across many files and offers the split before starting:

This touches 34 files with the same change. I can hand the edits to Codex — runs on your ChatGPT quota and keeps 34 file dumps out of this context — then review its diff here. Want me to?

What's inside

Skills — Claude picks these up automatically; you can also call them as /codex-bridge:<name>.

Subagents — Claude launches these itself, or you can name them.

Executables — on your PATH while the plugin is enabled, and usable straight from a shell.

Feature 1 — images

codex-imagegen "flat vector rocket icon, #2563EB on white, 2px stroke, minimal" ./rocket.png
codex-imagegen "change the sky to sunset orange, keep everything else identical" ./out.png --ref ./hero.png
codex-imagegen "settings gear, same style as the reference" ./settings.png --ref ./home.png

A custom size is only accepted if all of these hold: longest edge ≤ 3840px, both edges multiples of 16, long-to-short ratio ≤ 3:1, total pixels between 655,360 and 8,294,400. Square is fastest. For anything outside that — favicons, a 1200×630 OG card — generate a large master and resample locally.

The wrapper prints the path it actually wrote. Codex is instructed not to overwrite existing assets, so it occasionally saves out-v2.png; the wrapper detects that (and the ~/.codex/generated_images/<session>/ default location) and reports the real path.

What makes this better than pasting prompts into a chat window: the skills teach Claude to pull your actual palette out of tailwind.config.* or your design tokens, to reuse one style contract across a whole set so the assets match, to derive size variants with ImageMagick instead of regenerating, and to open the PNG and check it before telling you it's done.

Feature 2 — GPT-5 subagents

The point isn't "Claude can also call GPT-5". It's the division of labour: Claude scopes and verifies, Codex does the volume.

you ──▶ Claude          scopes the work, writes the brief, reviews the result
             │
             ▼
        codex-run ──▶ Codex CLI ──▶ GPT-5 / gpt-image-2
                                     (your ChatGPT quota)

Because Codex runs as a separate process, its intermediate output — the 40 files it read, the search results, the generated boilerplate — never enters your Claude context. You pay Claude tokens only for the brief and the review.

What gets delegated, per the codex-delegate rubric: mechanical edits across many files, bulk generation, exhaustive repo sweeps, all image work, and self-contained subtasks. What doesn't: architecture calls, ambiguous requirements, anything needing the conversation history, and small edits — round-tripping a two-line fix costs more, not less.

Every subagent follows the same contract: read-only by default, verify before reporting, and never claim success on unverified work. Only codex-implementer writes to your files, only after you've agreed, and only with a clean working tree so its changes land as a reviewable diff.

Straight from a shell:

codex-run -C . "which module owns rate limiting? name the file"
codex-run -C . -r "now show me the minimal patch"          # -r continues the session
git diff main...HEAD | codex-run -C . --timeout 1200 -      # pipe a diff in

How it works

Both wrappers shell out to codex exec, Codex's non-interactive mode.

Images run in a workspace-write sandbox scoped to the output directory:

codex exec -C <outdir> -s workspace-write --skip-git-repo-check [-i <ref>...] \
  "Use the \$imagegen image generation tool to ... save to ./<file>.png ..."

$imagegen is Codex's built-in image skill; it calls gpt-image-2 with your ChatGPT credentials and writes the PNG into the working directory. The wrapper verifies the file landed — falling back to $CODEX_HOME/generated_images/ if Codex saved it there instead — and prints the absolute path so Claude can read the image back.

Text tasks capture the final message with -o instead of scraping the event log, so subagents get a clean answer rather than a transcript:

codex exec -C <dir> -s read-only --skip-git-repo-check -o <tmpfile> "<task>"

Both fail fast and loudly: a missing codex binary or a logged-out session exits 1 with an actionable message rather than hanging. Timeouts exit 124. Both handle BSD and GNU stat, and degrade gracefully when timeout/gtimeout isn't installed.

Limitations

• Quota, not free. Image turns burn ChatGPT plan quota roughly 3–5× faster than text turns. Set OPENAI_API_KEY and Codex switches to API billing instead.
• Transparency takes two steps. The default path can't emit alpha, so the skills generate on a flat #00FF00 key and strip it with the remove_chroma_key.py helper Codex already ships. True native transparency needs Codex's CLI fallback plus an OPENAI_API_KEY; the skills surface that as a choice rather than switching silently.
• Slow. 1–4 minutes per image; a repo-wide Codex task can take 10+ minutes. Skills set long Bash timeouts accordingly.
• Codex is blind to your conversation. Every brief has to stand alone. This is exactly why the subagents verify before reporting.
• macOS and Linux. Not tested on Windows.

Troubleshooting

Contributing

Issues and PRs welcome.

claude plugin validate .
bash -n bin/codex-imagegen && bash -n bin/codex-run

Credits

• openai/codex — the Codex CLI and its $imagegen skill
• Codex CLI image generation write-up

License

MIT — see LICENSE.