chatgpt-image-mcp

An MCP server (and CLI) that generates images through the ChatGPT web backend and assembles them into full-bleed PowerPoint decks (brand-aware and reference-styled).

⚠️ Disclaimer

This project is provided for personal learning and research only. It reverse-engineers the ChatGPT web backend:

• Not affiliated with, endorsed by, or sponsored by OpenAI.
• Using it violates OpenAI’s Terms of Service. Your account may be rate-limited or permanently banned.
• Use a throwaway / secondary account — never your important one.

Provided as-is, with no warranty. You assume all risk. If you are not comfortable with these terms, do not use this software.

What it does

• Generates images from text prompts at a chosen aspect ratio (16:9, 1:1, 3:4, 9:16, or raw WxH).
• Auto-enhances prompts via your ChatGPT account’s text model before drawing (mirrors what the web UI does silently). Three slide styles: auto, slide, fintech.
• Builds full-bleed PowerPoint (.pptx) decks from a set of images.
• Branded decks: auto-detects brand colors from your logo, generates slides in those colors, and composites your logo onto every slide.
• Styled decks: matches the design style and palette of a reference image (without copying its text/content).
• Works as an MCP server across Claude Code, Codex, and Antigravity over stdio.
• Single-account, fully local auth — your token never leaves your machine.

Requirements

• uv — handles Python and dependencies (you do not need to install Python separately; uv fetches Python 3.12 automatically).
• git
• A ChatGPT account (use a secondary one — see disclaimer)

Install uv (one time per machine):

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

Install

One-click installer (recommended)

After cloning, run the installer — it installs uv if missing, syncs deps, walks you through login, and registers the MCP server:

git clone https://github.com/andyluu98/chatgpt-image-mcp.git
cd chatgpt-image-mcp

# Windows (PowerShell) — or right-click install.ps1 -> Run with PowerShell
powershell -ExecutionPolicy Bypass -File install.ps1

# macOS / Linux
bash install.sh

Manual install

git clone https://github.com/andyluu98/chatgpt-image-mcp.git
cd chatgpt-image-mcp
uv sync
uv run cgimg login
# A browser opens -> log into ChatGPT -> you land on a platform.openai.com page (it may say "Oops").
# Copy the FULL callback URL from the address bar, then run:
uv run cgimg login --callback ""

The repo is private — the cloning machine must be signed into a GitHub account with access (e.g. gh auth login, or git credentials for andyluu98).

Why login is two steps

Login uses OAuth with PKCE. Step 1 builds the authorization URL and stashes a one-time secret (the PKCE verifier) on disk. Step 2 exchanges the code in the callback URL for tokens — and that exchange needs the same verifier that step 1 generated. Splitting it into two commands lets the verifier persist between building the URL and redeeming the code, instead of being lost when the browser hands control back to you.

Register as an MCP server

The config is the same shape for Claude Code, Codex, and Antigravity:

{
  "mcpServers": {
    "chatgpt-image": {
      "command": "uv",
      "args": ["run", "cgimg-mcp"],
      "cwd": ""
    }
  }
}

Replace cwd with the absolute path where you cloned the repo (e.g. F:\\chatgpt-image-mcp on Windows — note the doubled backslashes in JSON). For Claude Code you can instead run:

claude mcp add chatgpt-image -s user -- uv run --directory "" cgimg-mcp

-s user makes the server available in every folder; --directory lets uv find the project from any working directory.

Restyle a building render (ref_mode="render")

Pass a 3D render (Blender, SketchUp, Lumion…) as ref_image with ref_mode="render" (aliases restyle, architecture). The model keeps massing, floors, openings, roof and camera, and changes only materials, lighting and landscaping. The prompt enhancer is skipped in this mode.

uv run cgimg gen "Indochine style, warm afternoon light" --ref-image render.png --ref-mode render --out out

MCP tools

Six tools are exposed by the server (src/cgimg/server.py):

For generate_slide_deck / branded_deck / styled_deck, each prompt is the content of one slide — pass raw slide content and the enhancer designs it.

Slide styles

When enhance=True, the prompt is expanded by your ChatGPT account’s text model before drawing. The style argument picks the design treatment:

Content completion (for slide / fintech): these styles complete your content into a full, information-rich slide — every main point gets a bold label plus a 2-line supporting description, sparse input is intelligently expanded into a sensible slide, and the prompt explicitly demands that all text be rendered in full (never dropped or abbreviated). It stays legible (not a wall of tiny text) and never fabricates fake statistics. See docs/styles.md for a full reference.

generate_image and the CLI gen accept style values auto, slide, and fintech. generate_slide_deck defaults to slide.

CLI usage

# 1. Log in (two-step, see Install above)
uv run cgimg login
uv run cgimg login --callback ""

# 2. Generate image(s)  (auto-enhance is ON by default; add --no-enhance to send the prompt as-is)
uv run cgimg gen "a serene mountain lake at dawn" --aspect 16:9 --n 1 --out out
uv run cgimg gen "AI agents for customer support" --style slide      # clean editorial slide
uv run cgimg gen "real-time fraud detection" --style fintech         # light-blue dashboard slide
uv run cgimg gen "ai agent" --aspect 1:1 --no-enhance                # send prompt verbatim

# 3. Build a deck from existing images
uv run cgimg ppt img1.png img2.png --out deck.pptx --aspect 16:9

# 4. Branded deck — slides in your brand colors with your logo composited on each
uv run cgimg branded logo.png \
  --prompts "What is RAG?" "RAG pipeline" "Benefits" \
  --out deck.pptx --position top-left --scale 0.15

# 5. Styled deck — match a reference image's design (its text/content is NOT copied)
uv run cgimg styled reference-slide.png \
  --prompts "Intro" "How it works" "Pricing" \
  --out deck.pptx

branded and styled also accept --aspect and --out-dir (default out). They print the deck path, the detected brand_colors, and each generated image path.

Aspect ratios

ChatGPT honors the ratio, not exact pixels — it normalizes to its own native dimensions (so 16:9 yields roughly 1672x941, not exactly 1920x1080). This is expected.

PPTX output is full-bleed: the image fills the slide edge to edge, so the image aspect should match the deck aspect to avoid cropping or letterboxing.

How it works

• Vendors chatgpt2api’s proven OAuth + image backend (under src/cgimg/_vendor/).
• Stores a single account token locally at %APPDATA%\cgimg\auth.json (Windows) or ~/.config/cgimg/auth.json (Linux/macOS). It is never committed.
• Auto-refreshes the access token via the stored refresh_token when it expires.

Examples

Showcase slides generated by this server live in examples/sample-slides/ — a mix of slide/fintech styles, dense multi-task slides, custom layouts, and tables.

Troubleshooting

Attribution & License

This project vendors code from:

• chatgpt2api — Copyright © 2026 kunkun, MIT License. Powers the OAuth + image backend. Vendored under src/cgimg/_vendor/.

Vendored files retain their original behavior. See NOTICE for details.