AI Image Preset Editor — MCP app for ChatGPT

A visual editor inside ChatGPT that builds deterministic AI image prompts. Pick a size, a layout, place your texts, choose style, palette and restrictions — and watch the prompt compile live. Then just say "generate the image" and ChatGPT renders it. Save the result as a preset and reuse it forever: "open my promo dark preset".

Same preset in → same prompt out. Always. To be precise: the prompt is deterministic; image models are still stochastic. What you get is a structured, reproducible, versionable prompt instead of a hand-written one that's different every time.

What it looks like

You: "I want to compose an image" → the editor opens inside the chat:

Size — aspect ratio chips (1:1, 4:5, 9:16, 16:9, 3:2) or custom pixels
Layout — 9 visual patterns: full bleed, vertical/horizontal split, diagonal split, 2×2 grid, 3 columns, hero + bar, frame, overlay
Elements — your exact texts (rendered verbatim) and subjects, each placed on a 3×3 position grid
Style — art style, color palette, typography, lighting
Restrictions — no watermarks, no extra text, no people… plus your own

The compiled prompt updates live with every click. Then:

Generate 🎨 — sends the prompt to ChatGPT, which generates the image natively
Save — "save it as promo dark" works too, from the conversation
Later: "bring up the preset I used yesterday" → the editor reopens pre-loaded

Quickstart

1. Run your instance

This is a self-hosted, single-user server — your presets live in one SQLite file on your instance. No accounts, no tracking, no cloud database.

Docker (any host):

docker build -t preset-editor .
docker run -p 8080:8080 -v preset-data:/data preset-editor

Google Cloud Run (free tier friendly):

gcloud run deploy preset-editor --source . --allow-unauthenticated --region us-central1

Note: on Cloud Run the filesystem is ephemeral — saved presets reset on redeploys/restarts unless you mount a volume (Cloud Run volume mounts or any host with a persistent disk). For durable presets prefer a small VPS, Fly.io, Railway with a volume, or your own machine + a tunnel (e.g. cloudflared, ngrok).

Local (development):

python -m venv .venv && source .venv/bin/activate
pip install .[dev]
python -m server.app   # → http://localhost:8080/mcp

Your MCP endpoint is https://YOUR-HOST/mcp — ChatGPT needs a publicly reachable HTTPS URL.

2. Connect it to ChatGPT

ChatGPT → Settings → Apps & Connectors → Advanced → Developer mode (enable it)
Create a new connector
Name: Preset Editor · MCP Server URL: https://YOUR-HOST/mcp · Auth: None
Save. In a new chat, enable the connector (＋ → your connector if needed)

3. Use it

| You say | What happens | |---|---| | "I want to compose an image" | The visual editor opens in the chat | | (click around) | The prompt compiles live in the panel | | "Generate the image" / Generate 🎨 | ChatGPT generates with the exact compiled prompt | | "Save it as promo dark" / Save | Preset stored on your instance | | "List my presets" | Names + when you last used each | | "Open the one I used yesterday" | Editor reopens pre-loaded |

Why deterministic prompts?

Hand-written prompts drift: every retype changes words, order, emphasis — and your brand look with them. Here a preset is data (JSON), and the prompt is compiled from it with fixed templates, fixed order, zero randomness, zero LLM:

{
  "size": { "width": 1080, "height": 1350, "aspect_label": "4:5" },
  "layout": "diagonal",
  "elements": [
    { "kind": "text", "content": "50% OFF", "zone": "top-right" },
    { "kind": "subject", "content": "a steaming artisan pizza", "zone": "center" }
  ],
  "style": { "art_style": "photorealistic", "palette": ["#1a1a2e", "#e94560"] },
  "restrictions": ["no_watermarks", "no_extra_text"]
}

compiles — always — to:

Output format: 1080x1350 pixels (4:5).

Layout pattern: Diagonal split: the canvas is divided into two contrasting triangular zones by a diagonal line.

Element placement:
  - "50% OFF" (text) -> positioned at top-right
  - a steaming artisan pizza (subject) -> positioned at center

Visual style — art style: photorealistic, color palette: #1a1a2e #e94560.

Do NOT include: watermarks, any text other than the quoted texts.

Follow the layout, element placement, palette and restrictions exactly as specified. Render every quoted text verbatim, with no spelling changes or additions.

Two guarantees back this up:

Golden tests: every fixture in core/tests/golden/ must compile to its exact expected output, byte for byte — in both the Python compiler and its JS port (the widget's live preview). CI fails if they diverge by one character.
Versioned compilers: each preset records its compiler_version. Compilers are frozen on release — future improvements ship as v2, v3… and old presets keep producing their original prompt forever.

MCP tools

| Tool | Purpose | |---|---| | open_preset_editor | Opens the widget (optionally pre-loaded with a saved preset) | | compile_preset | Preset JSON → deterministic prompt | | save_preset | Save/update by name (also used by the widget's Save button) | | list_presets | Filter by name, sort by last_used / created / name | | get_preset | Fetch a preset + its compiled prompt | | delete_preset | Remove one | | get_preset_options | Reference of all valid values (layouts, zones, styles…) |

Development

pip install .[dev]
pytest                              # Python: compiler goldens + storage + MCP e2e
node widget/test/golden.test.js     # JS port == Python compiler, byte for byte
python widget/dev/serve.py          # widget UI at http://127.0.0.1:8001 (mock host)

The widget dev server injects a fake window.openai, so you can iterate on the UI in a normal browser — no deploy, no ChatGPT.

After an intentional compiler change (a new version!): python -m core.tests.regen_goldens and review the diff.

core/      schema + constants + versioned compilers + golden fixtures
server/    FastMCP server: tools, SQLite storage, widget assembly
widget/    the editor: src/ (inlined into ONE html), dev/ (mock), test/ (parity)

Roadmap

Export / import presets as portable JSON
Community preset gallery
Standalone web playground (no ChatGPT needed)
Multi-user mode with OAuth — contributions welcome

License

MIT

MCP Servers