image-gen

MCP image generation for AI coding agents — no OpenAI API key

Drives your own logged-in ChatGPT session in a dedicated browser profile,
downloads the PNG, and drops it into your project.

Why this exists

Most agent image tools need an OpenAI API key and burn credits. image-gen borrows the ChatGPT plan you already pay for (Free / Plus / Pro), runs generation in a real browser, and returns a local file path your agent can embed immediately.

Built for Cursor, Claude Code, Claude Desktop, and any MCP client.

The reliable way to use it is the bundled skill. Register the MCP, copy skills/chatgpt-image-gen/ into your agent, then invoke /chatgpt-image-gen (or @chatgpt-image-gen) and describe the asset. The agent drives your ChatGPT session and saves a real PNG into the project. Setup is below.

Features

• One-time login into a dedicated Edge auth profile (does not touch your daily browser)
• Attach-first CDP — reconnects if the browser is already open; never force-kills your normal Edge
• Auto-clicks ChatGPT’s Welcome back / Choose an account picker
• Session backup via chatgpt-storage.json (cookies restored if needed)
• Reference-image uploads for variations / style matching
• Auto-dismisses ChatGPT’s “image was already uploaded” / duplicate-file modal
• Transparent-background prompt helper for icons and UI marks
• Transparency anti-hallucination — alpha report + checkerboard .preview.png so agents don’t loop regenerating over ChatGPT’s soft bottom shadow
• Thread-safe image detection (ignores older images already in the chat)
• Parallel generate_image calls are queued (in-process + file lock)

Requirements

Quick start (5 minutes)

1. Install

git clone https://github.com/nothariharan/image-gen.git
cd image-gen
npm install

2. One-time ChatGPT login

npm run login

What happens:

1. A dedicated Edge window opens (edge-auth-profile/) on port 9222
2. Your normal/daily Edge is left alone
3. Log into ChatGPT fully in that window (Google / password / 2FA)
4. When the chat composer appears, the script saves chatgpt-storage.json and exits

Do not delete these after login:

• edge-auth-profile/ — browser profile with your session
• chatgpt-storage.json — cookie backup (gitignored)

3. Verify

npm run test-connection
# → Connected! The browser has N open context(s).

4. Register the MCP with your agent

Use the absolute path to mcp-server.mjs on your machine.

Cursor

Open ~/.cursor/mcp.json (Windows: %USERPROFILE%\.cursor\mcp.json) and add:

{
  "mcpServers": {
    "image-gen": {
      "command": "node",
      "args": ["C:/Users/YOU/image-gen/mcp-server.mjs"]
    }
  }
}

Reload Cursor (or toggle the MCP off/on). The tool appears as generate_image on server image-gen / user-image-gen.

Claude Code

claude mcp add -s user image-gen node /absolute/path/to/image-gen/mcp-server.mjs

Claude Desktop

Edit your Claude Desktop MCP config and add the same command / args block as Cursor.

5. Install the skill and invoke it

Agents often do not pick this MCP on their own. They see an image request and call a built-in image tool instead. The skill is the switch: you invoke it, and the agent has to call generate_image.

The skill ships in this repo:

skills/chatgpt-image-gen/SKILL.md

Copy that folder into your agent's skills directory (once):

Cursor (Windows PowerShell) — run this from the cloned repo:

New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.cursor\skills\chatgpt-image-gen" | Out-Null
Copy-Item .\skills\chatgpt-image-gen\SKILL.md "$env:USERPROFILE\.cursor\skills\chatgpt-image-gen\SKILL.md"

Claude Code (Windows PowerShell) — same folder, under .claude:

New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\skills\chatgpt-image-gen" | Out-Null
Copy-Item .\skills\chatgpt-image-gen\SKILL.md "$env:USERPROFILE\.claude\skills\chatgpt-image-gen\SKILL.md"

Cursor and Claude Code (macOS / Linux):

mkdir -p ~/.cursor/skills/chatgpt-image-gen ~/.claude/skills/chatgpt-image-gen
cp skills/chatgpt-image-gen/SKILL.md ~/.cursor/skills/chatgpt-image-gen/SKILL.md
cp skills/chatgpt-image-gen/SKILL.md ~/.claude/skills/chatgpt-image-gen/SKILL.md

Start a new chat, then invoke the skill and describe the asset:

/chatgpt-image-gen app icon, flat, dark blue, transparent background, save it in public/icons

@chatgpt-image-gen works the same if your client uses @ mentions. You can also attach the skill in the composer the way you attach any other skill.

That is the whole loop after setup:

1. You invoke the skill and say what you want.
2. The agent calls generate_image on the image-gen MCP (never its built-in image tool).
3. ChatGPT draws it in your logged-in browser.
4. A PNG lands in public/ or assets/, and the agent embeds that file.

Everyday use

Once login and the skill are in place, invoke /chatgpt-image-gen and your agent calls:

generate_image

Parallel calls are queued (important)

ChatGPT is a single shared browser tab. If two generate_image jobs type at once, prompts interleave into garbage text.

v2.4+ serializes every job with an in-process queue + cross-process .generate.lock file. Parallel MCP calls wait their turn instead of colliding. You may still call multiple times from the agent — they run one after another automatically.

If a job crashes hard and leaves a stuck lock, delete image-gen/.generate.lock (or wait until it goes stale, ~6 minutes).

Example calls

{
  "prompt": "rainy city street at night, neon reflections, cinematic, wide shot",
  "filename": "hero",
  "output_dir": "./public"
}

{
  "prompt": "minimalist calendar icon, thin lines, dark blue",
  "filename": "icon-calendar",
  "output_dir": "./public/icons",
  "transparent_background": true
}

{
  "prompt": "same character but angrier, keep the art style",
  "filename": "mascot-angry",
  "output_dir": "./public",
  "reference_images": ["C:/path/to/project/public/mascot.png"]
}

Returns: absolute path of the saved PNG.

How auth persistence works

npm run login  →  edge-auth-profile + chatgpt-storage.json
        ↓
generate_image
        ↓
  Is port 9222 up? ──yes──► attach (no browser kill)
        │
        no
        ↓
  Launch ONLY edge-auth-profile on :9222
        ↓
  Restore cookies from chatgpt-storage.json if needed
        ↓
  Auto-click Welcome-back account picker if shown
        ↓
  Generate + download PNG + refresh storage file

When might ChatGPT still ask you to log in again?

OpenAI does not publish an exact cookie TTL. In practice, sessions often last weeks to months if the auth profile is left alone. Re-login is usually triggered by things like:

Day-to-day use of this MCP should not log you out by itself anymore (it no longer kills your daily browser or mixes profiles).

Environment variables

Example (Windows PowerShell):

$env:IMAGE_GEN_CHATGPT_EMAIL = "[email protected]"
node mcp-server.mjs

Or in Cursor mcp.json:

{
  "mcpServers": {
    "image-gen": {
      "command": "node",
      "args": ["C:/Users/YOU/image-gen/mcp-server.mjs"],
      "env": {
        "IMAGE_GEN_CHATGPT_EMAIL": "[email protected]"
      }
    }
  }
}

How generation works (technical)

Playwright attaches over CDP, finds/opens a chatgpt.com tab, types the prompt, and waits for:

img[alt^="Generated image"]

Primary path waits until the src stops changing for ~3 seconds (preview → final swap), then fetches the signed Estuary URL from inside the page.

Fallback chain

1. Stable img[alt^="Generated image"] → in-page fetch
2. Largest image/* network response during generation (≥ 80 KB)
3. Newest large <img> in the DOM

Conversation streams are parsed only to detect refusals early (3‑minute hard timeout otherwise).

Tab / thread behavior

• Reuses the first open chatgpt.com tab; leaves it open afterward
• Keep the tab open → generations stay in one thread (good for iteration)
• Close the tab → next call starts a fresh conversation
• Want a specific thread → open that URL in the auth-profile browser first

Collision guard: snapshots existing generated-image srcs before sending the prompt, then only accepts new ones.

Scripts

macOS / Linux

There is no .bat helper, but the flow is the same: dedicated profile + remote debugging + npm run login.

# Example: Chrome with a dedicated profile
google-chrome \
  --remote-debugging-port=9222 \
  --remote-allow-origins=* \
  --user-data-dir="$PWD/edge-auth-profile" \
  https://chatgpt.com

Then in another terminal:

npm run login   # if you still need to complete/save login
npm start

Troubleshooting

Security & privacy

• Session files are local secrets. They are gitignored. Never commit or share chatgpt-storage.json or edge-auth-profile/.
• The MCP talks to ChatGPT through your browser session — treat the machine like it’s logged into ChatGPT.
• This project is not affiliated with OpenAI. Automating ChatGPT may conflict with OpenAI’s terms depending on use; for production/commercial scale, prefer the official Images API.

Limitations

• Slow — typically 15–60s per image (3 min hard cap)

• One image at a time — jobs are queued and run serially

• DOM can change — selectors may need updates after big ChatGPT UI changes

• Account rate limits apply (especially Free)

• Not a drop-in replacement for high-volume API generation

Project layout

image-gen/
├── mcp-server.mjs       # MCP entry (stdio) + transparency report response
├── generate.mjs         # CDP attach, login, generation, download
├── transparency.mjs     # Alpha inspect + checkerboard .preview.png
├── queue.mjs            # Serialize parallel generate_image calls
├── login-once.mjs       # One-time auth setup
├── setup-session.mjs    # Connection smoke test
├── launch-edge.bat      # Windows auth-profile launcher
├── skills/
│   └── chatgpt-image-gen/
│       └── SKILL.md     # Invoke with /chatgpt-image-gen so the agent uses this MCP
├── edge-auth-profile/   # Created locally — gitignored
├── chatgpt-storage.json # Created locally — gitignored
└── assets/

License

MIT — not affiliated with OpenAI or Anthropic.