pi-imagegen

A Pi package for generating images with your existing OpenAI/Codex subscription session.

pi-imagegen adds an agent-callable imagegen tool, a /img command namespace, and a local browser studio for visual image workflows.

What it does

• Generates images through Pi's existing openai-codex OAuth login.
• Uses the Codex Responses backend with native gpt-image-2 image generation.
• Saves images and sidecar metadata locally.
• Supports batches, style presets, reference images, and sketch references.
• Provides a browser-based studio for browsing, comparing, rerunning, varying, and referencing images.

Install

Install the published npm package:

pi install npm:pi-imagegen

npm package: https://www.npmjs.com/package/pi-imagegen

For local development:

git clone https://github.com/Jon-Vii/pi-imagegen.git
cd pi-imagegen
pi install .

Then reload Pi:

/reload

You also need to be logged into Pi's OpenAI/Codex provider:

/login

Select the ChatGPT/Codex option that provides the openai-codex provider.

Commands

/img studio
/img gen [--thinking off|minimal|low|medium|high] [--style name] <prompt>
/img batch <count> [--thinking off|minimal|low|medium|high] [--style name] <prompt>
/img styles
/img list [count]
/img open [latest|number|path]
/img reveal [latest|number|path]
/img path [latest|number|path]
/img info [latest|number|path]

Examples:

/img gen tiny blue ceramic fish on white background
/img gen --thinking off --style poster a cinematic expedition poster for a lava cavern
/img batch 4 --style wallpaper a quiet mountain observatory at sunrise
/img studio

Studio

Run:

/img studio

The studio opens a local browser UI served from 127.0.0.1.

It supports:

• image history wall
• grouped batch/contact-sheet view
• modal preview with rerun/vary/reference actions
• prompt composer with style, aspect, quality, thinking, and count controls
• real image references sent as input_image content
• sketch references via a simple drawing canvas

A typical loop:

Draw or pick reference → Generate 4 → inspect → Vary or Rerun → keep exploring

Agent tool

The package also registers a model-facing tool:

imagegen

It can generate an image and return both a saved file and an inline image attachment. It supports options such as:

• prompt
• size
• quality
• background
• outputFormat
• thinking
• referencePaths
• outputPath

How it works

This package does not use the public OpenAI image API or a separate API key.

It uses Pi's existing openai-codex OAuth token and calls:

https://chatgpt.com/backend-api/codex/responses

with the native Responses image generation tool:

{
  "type": "image_generation",
  "model": "gpt-image-2"
}

Generated image results are received from streamed SSE events and saved locally.

Files and metadata

By default, generated images are saved under:

~/.pi/agent/generated-images/

Each image gets a JSON sidecar with prompt, model, path, format, reference, batch, and generation metadata.

Batches are saved under:

~/.pi/agent/generated-images/batches/

Sketch references are saved under:

~/.pi/agent/generated-images/sketches/

Notes

This package relies on Pi internals and the Codex Responses backend. It is intended for personal/local Pi workflows and may need updates if the upstream backend changes.

License

MIT