c2c-img — built-in image generation for Claude Code, powered by Codex

English | 简体中文

If you have both Claude Code and Codex (ChatGPT), this plugin lets Claude generate and edit images with Codex's built-in image_gen tool. It runs on your ChatGPT login, so you don't need an OPENAI_API_KEY.

It is designed to feel native. You say "draw me a…", and a moment later the image appears in the conversation. Say "make it brighter" or "now a portrait version", and Claude edits the last image. Claude writes the prompt, checks the result itself, and never shows you the plumbing.

Chinese poster, text rendered exactly	Edit: change dawn to an aurora night, and nothing else	Transparent sticker (-t -s 512x512 -f webp)
	→

Install

Requirements:

• The Codex CLI, from the ChatGPT desktop app or npm i -g @openai/codex, signed in with codex login using a ChatGPT account.
• Python 3.9 or later.

As a Claude Code plugin (recommended):

/plugin marketplace add actionow-ai/c2c_img
/plugin install c2c-img@c2c-img

As a plain skill, for Claude Code and other agents that support Agent Skills:

npx skills add actionow-ai/c2c_img -g -a claude-code

Or copy skills/c2c-img/ into ~/.claude/skills/.

Check your setup (this generates nothing):

python3 skills/c2c-img/scripts/c2c_img.py --doctor

pip install pillow is optional. It adds webp output, and on Linux and Windows it also adds --size and jpg support. On macOS, the built-in sips already handles resizing, png and jpg.

Use it

Just ask: "Draw a cozy isometric reading nook with a sleeping cat", "make a 1200×630 OG image for this blog post", "remove the background from ./logo.png", or "做一张中秋海报". Claude picks sensible defaults (aspect ratio, style, where to save), generates the image, checks it, and shows it to you.

Under the hood, Claude calls the script:

Without -o, images are saved to ~/Pictures/c2c_img/<time>-<slug>.png. Existing files are never overwritten: a new version is saved as -v2, -v3, and so on.

How it works

Claude ─Bash─▶ c2c_img.py ─▶ 127.0.0.1 + token ─▶ daemon ─▶ codex app-server (kept warm)
                          └─ fallback ─▶ codex exec --ephemeral (lean mode), one process per image

• Daemon engine (default): keeps codex app-server warm between calls, cutting the median overhead per call from 13.9 s to 6.6 s. It reads the saved path directly from imageGeneration events. It gets structured errors: quota with its reset time, capacity (it retries on another model), and refusals. It exits after 15 idle minutes, and it's replaced automatically when Codex or the script is upgraded.
• Exec engine (fallback): runs codex exec with the user's config ignored and 17 unneeded tools disabled, including shell. It stops Codex as soon as the image file is complete.
• Both engines run Codex in an empty temp directory with a read-only sandbox. They move (not copy) the result out of ~/.codex/generated_images.

A typical new image takes about 40–50 s. Most of that is the image model itself, and it varies with server load.

Platforms and tests

macOS has been tested with a real Codex install. Linux and Windows are supported through process groups and taskkill, cross-platform file locks, and UTF-8 stdio. CI covers macOS, Ubuntu and Windows with Python 3.9 and 3.12.

python3 tests/test_offline.py

python3 tests/test_engines.py

• test_offline.py: unit tests and post-processing.
• test_engines.py: end-to-end tests of both engines against tests/fake_codex.py. No Codex login or quota is needed.

Limitations

• It relies on Codex internals: the app-server protocol (marked experimental), exec --json events, and the generated_images/ layout. The two engines back each other up. After a Codex upgrade, run --doctor.
• The image tool has no size parameter. The prompt sets the aspect ratio, and -s post-processes to exact pixels.
• Each image uses your ChatGPT/Codex quota.

License

MIT © 2026 actionow