Node.js library and CLI for sending image-generation requests to Codex's private ChatGPT-authenticated backend path.

WARNING: This is not a supported public API integration. It depends on private Codex request behavior that may change without notice.

• Reuses local Codex ChatGPT auth from ~/.codex/auth.json
• Reads ~/.codex/installation_id when available
• Sends a POST request to https://chatgpt.com/backend-api/codex/responses
• Requests the built-in image_generation tool with output_format: png
• Parses streamed SSE output and saves the resulting PNG
• Supports dry-run and sanitized debug dumps with request/response metadata minimization
• Also supports a codex exec fallback provider that verifies real PNG output from ~/.codex/generated_images/

• Node.js 20+
• Existing local Codex ChatGPT login state
• A Codex/ChatGPT account that is entitled to image generation on the private backend

• Node.js 20+ (for CLI and Node.js library)
• Python 3.10+ (for Python SDK)
• Existing local Codex ChatGPT login state (~/.codex/auth.json)
• A Codex/ChatGPT account entitled to image generation on the private backend

npm install -g god-tibo-imagen

After installation, the gti command is available globally:

gti --version
gti --help

npm install god-tibo-imagen

import { createProvider, resolveConfig } from 'god-tibo-imagen';

pip install god-tibo-imagen

from gti import Client

npm test
npm run check
gti --prompt "flat blue square icon" --output ./out/blue-square.png

You can provide existing images as additional context alongside your text prompt. Images are embedded as base64 data URLs and sent with the request. Use --image multiple times for multiple images.

# single image
gti --prompt "Make this cat wear a hat" --image ./cat.png --output ./cat-hat.png

# multiple images
gti --prompt "Combine these two styles" --image ./style-a.png --image ./style-b.png --output ./combined.png

Supported formats: png, jpg/jpeg, gif, webp.

Pass --size <value> to control the output image dimensions. Supported values match the gpt-image-2 spec:

gti --prompt "a sunset over mountains" --size 1536x1024 --output ./sunset.png

Supported sizes:

• auto (model decides)
• 1024x1024, 2048x2048 (square)
• 1536x1024, 2048x1152, 3840x2160 (landscape)
• 1024x1536, 2160x3840 (portrait)

The --size flag is forwarded to the image_generation tool config and is honored by the private Codex backend. The codex-cli provider does not support --size; direct codex-cli use and auto fallback fail fast rather than silently ignoring requested dimensions.

# direct private HTTP path
gti --provider private-codex --prompt "flat blue square icon" --output ./out.png

# borrow the Hermes-style codex exec workaround
gti --provider codex-cli --prompt "flat blue square icon" --output ./out.png

# try private HTTP first, then fall back to codex-cli
gti --provider auto --prompt "flat blue square icon" --output ./out.png

gti --prompt "flat blue square icon" --dry-run

npm run smoke:live -- "Generate a tiny flat blue square icon." ./smoke-output.png

import { createProvider, resolveConfig, loadCodexSession, validateCodexSession } from 'god-tibo-imagen';

const config = resolveConfig({ provider: 'private-codex' });
const provider = createProvider(config);

const result = await provider.generateImage({
  prompt: 'flat blue square icon',
  model: 'gpt-5.4',
  outputPath: './out.png',
  dryRun: false,
  debug: false
});

console.log(result.savedPath);

You can also pass existing images as input:

// single image
const result = await provider.generateImage({
  prompt: 'Make this cat wear a hat',
  model: 'gpt-5.4',
  outputPath: './cat-hat.png',
  images: ['data:image/png;base64,iVBORw0KGgo...']
});

// with output size
const result = await provider.generateImage({
  prompt: 'a sunset over mountains',
  model: 'gpt-5.4',
  outputPath: './sunset.png',
  size: '1536x1024'
});

// multiple images
const result = await provider.generateImage({
  prompt: 'Combine these two styles',
  model: 'gpt-5.4',
  outputPath: './combined.png',
  images: [
    'data:image/png;base64,abc...',
    'data:image/png;base64,def...'
  ]
});

from gti import Client

client = Client(provider="private-codex")
result = client.generate_image(
    prompt="flat blue square icon",
    model="gpt-5.4",
    output_path="./out.png"
)
print(result.saved_path)

# with output size
result = client.generate_image(
    prompt="a sunset over mountains",
    model="gpt-5.4",
    output_path="./sunset.png",
    size="1536x1024"
)
print(result.saved_path)

You can also pass existing images as input:

# single image
result = client.generate_image(
    prompt="Make this cat wear a hat",
    model="gpt-5.4",
    output_path="./cat-hat.png",
    image_paths="./cat.png"
)

# multiple images
result = client.generate_image(
    prompt="Combine these two styles",
    model="gpt-5.4",
    output_path="./combined.png",
    image_paths=["./style-a.png", "./style-b.png"]
)

gti --prompt "flat blue square icon" --output ./out.png

import { createProvider, resolveConfig } from 'god-tibo-imagen';

const config = resolveConfig({ provider: 'private-codex' });
const provider = createProvider(config);

const result = await provider.generateImage({
  prompt: 'flat blue square icon',
  model: 'gpt-5.4',
  outputPath: './out.png',
});

console.log(result.savedPath);

from gti import Client

client = Client(provider="private-codex")
result = client.generate_image(
    prompt="flat blue square icon",
    model="gpt-5.4",
    output_path="./out.png"
)
print(result.saved_path)

With image inputs:

result = client.generate_image(
    prompt="Make this cat wear a hat",
    model="gpt-5.4",
    output_path="./cat-hat.png",
    image_paths="./cat.png"
)
print(result.saved_path)

For users who want to invoke god-tibo-imagen from within any coding agent that supports the Agent Skills format (Claude Code, Codex, Cursor, OpenCode, Continue, Gemini CLI, etc.), a portable skill is provided in skills/god-tibo-imagen/.

• Python 3.10+
• pip install god-tibo-imagen (the import name is gti, not god_tibo_imagen)
• Local Codex auth in ~/.codex/auth.json with auth_mode = chatgpt

# Dry run
python skills/god-tibo-imagen/scripts/wrapper.py --prompt "flat blue square icon" --output ./test.png --dry-run

# Live generation
python skills/god-tibo-imagen/scripts/wrapper.py --prompt "flat blue square icon" --output ./out.png

# With image inputs
python skills/god-tibo-imagen/scripts/wrapper.py --prompt "Make this cat wear a hat" --image ./cat.png --output ./cat-hat.png

The skill definition is at skills/god-tibo-imagen/SKILL.md. Install it by copying or symlinking the skills/god-tibo-imagen/ directory into your agent's skills path:

• Claude Code: ~/.claude/skills/god-tibo-imagen/
• Codex: ~/.codex/skills/god-tibo-imagen/ (or project-local .agents/skills/god-tibo-imagen/)
• OpenCode: ~/.config/opencode/skills/god-tibo-imagen/
• Cursor / Continue / Gemini CLI / Kiro: .agents/skills/god-tibo-imagen/ (project)

Or use the Vercel skills CLI:

npx skills add NomaDamas/god-tibo-imagen --skill god-tibo-imagen

See skills/god-tibo-imagen/README.md for detailed installation per agent.

• src/auth/loadCodexSession.js — reads Codex auth state
• src/auth/validateSession.js — validates required private-backend fields
• src/codex/buildResponsesRequest.js — builds the /responses request
• src/codex/streamResponsesSse.js — parses SSE events
• src/codex/extractImageGeneration.js — finds image_generation_call
• src/providers/privateCodexProvider.js — live request/response orchestration
• src/providers/codexCliProvider.js — Hermes-style codex exec fallback with file verification
• src/providers/createProvider.js — provider selection and auto fallback
• src/cli/generate.js — CLI entry point

• This MVP supports the file-backed ~/.codex/auth.json path.
• If your Codex install stores auth only in a keyring and does not materialize auth.json, this MVP will not discover it yet.
• Debug dumps redact bearer tokens, account/session identifiers, installation IDs, cookies, and image payload base64, and store a minimized response summary instead of the raw response body.
• The architecture now supports both the direct private HTTP client and a Hermes-style codex exec fallback, while keeping the provider seam open for future app-server integration.