Imagegen Bridge

Use image generation from your Codex subscription—no OpenAI API key required.

Quick start · Commands · Examples · SDKs and skill · Dashboard and API · Documentation

Imagegen Bridge connects to the Codex OAuth session from your own subscription and exposes its image generation through a local CLI, dashboard, and HTTP API. You keep using your existing Codex login instead of configuring or paying for a separate OpenAI API key.

Quick start

You need Rust 1.94, a working codex executable, and an existing Codex login.

cargo install imagegen-bridge-cli
codex login
imagegen-bridge setup
imagegen-bridge doctor
imagegen-bridge generate \
  "A red paper fox on a charcoal background" \
  --output first-image.png \
  --preview

Open the local generation dashboard with:

imagegen-bridge dashboard

setup previews every change before applying it. Neither setup nor doctor spends an image generation unless you explicitly request and confirm a live probe.

Main commands

Run imagegen-bridge <command> --help for every option. The complete CLI reference is in docs/cli.md.

Common workflows

Generate and save an image

imagegen-bridge generate \
  "A studio portrait with deep red hair" \
  --size 1024x1536 \
  --quality high \
  --format png \
  --output portraits/red-hair.png \
  --metadata sidecar \
  --preview

Use --output-dir for generated filenames, --collision suffix to preserve an existing file, --open for the system image viewer, and --metadata embedded to store generation metadata inside the image. Metadata is disabled by default.

Generate multiple images

imagegen-bridge generate \
  "Four variations of a stone bridge in fog" \
  --count 4 \
  --batch-execution parallel \
  --output-dir batches/bridges

If a transport supports only one image per request, the bridge fans out bounded provider calls and preserves the requested order.

Edit with references

imagegen-bridge edit \
  "Change the jacket to dark blue" \
  --image ./portrait.png \
  --reference ./palette.png \
  --response-format artifact

Create transparent output

imagegen-bridge generate \
  "A small red fox mascot, full body" \
  --background transparent \
  --transparency auto \
  --format png \
  --output mascots/fox.png

auto uses native alpha when the provider supports it and otherwise applies the local chroma-key pipeline. Existing keyed images can be processed without a provider call:

imagegen-bridge background remove keyed-input.png \
  --output transparent.png \
  --key auto

Reuse settings and threads

imagegen-bridge preset create portrait-high --from request.json
imagegen-bridge generate "A red-haired woman" --preset portrait-high

imagegen-bridge generate \
  "Create the first character sheet" \
  --session-key character-design
imagegen-bridge generate \
  "Keep the character and show a side view" \
  --session-key character-design

Presets never retain image bytes, masks, reference images, or idempotency keys.

Use from code or agents

Install the agent skill with:

npx skills add Crimsab/imagegen-bridge \
  --skill generate-images-with-bridge

The SDKs and skill discover the same live provider capabilities as the CLI. See docs/sdks.md for examples.

Dashboard and API

imagegen-bridge dashboard opens the embedded UI and starts the local service when needed. The dashboard supports generation, edits, reference images, advanced controls, presets, asynchronous jobs, history, previews, metadata, and provider diagnostics. It is embedded in the Rust binary: no Node runtime or separate web server is required.

To run it manually:

imagegen-bridge serve --bind 127.0.0.1:8787

Open http://127.0.0.1:8787/dashboard, or send a native request:

curl --fail --silent --show-error \
  -H 'Content-Type: application/json' \
  -d '{"operation":"generate","prompt":"A small stone bridge in fog"}' \
  http://127.0.0.1:8787/v1/images

OpenAI-familiar generation and multipart edit routes are also available at /v1/images/generations and /v1/images/edits. Durable jobs use /v1/jobs and persist artifact-backed results across restarts. See docs/api.md for authentication, streaming, presets, job history, diagnostics, metrics, errors, and the full route contract.

Providers and models

codex-responses is the built-in Codex path and can route gpt-image-2, gpt-image-1.5, gpt-image-1, and gpt-image-1-mini. It authenticates with the existing Codex/ChatGPT OAuth session and the Codex Responses backend. It does not read OPENAI_API_KEY and is separate from the official OpenAI Platform API-key provider. Inspect its live capabilities with:

imagegen-bridge providers capabilities \
  --provider codex-responses \
  --json

imagegen-bridge generate \
  --model gpt-image-2 \
  "A translucent red glass sculpture"

Requests are checked against the selected provider before generation. Strict mode rejects unsupported combinations; --compatibility normalize applies only changes reported back in the response. Masks are currently unsupported by both Codex transports.

codex-app-server remains available as a fallback. Its upstream turn may occasionally finish without emitting an image item; the bridge treats that as an error and records content-safe structured counts of observed item types and statuses for diagnosis. An API-key-backed official OpenAI provider is reserved in configuration but is not implemented and never shares Codex OAuth handling.

Advanced controls

The native contract supports, where the provider allows them:

• size, aspect ratio, resolution, quality, format, compression, and background;
• multiple outputs with sequential or bounded parallel execution;
• negative-prompt, moderation, revised-prompt, and partial-image policies;
• input fidelity, reference images, edits, and reusable sessions;
• native or emulated transparency with alpha validation and despill;
• atomic output files, sidecar or embedded metadata, timings, and checksums;
• ordered provider/model fallbacks with per-attempt traces;
• strict, normalized, fail-fast, and best-effort request behavior.

Provider fallback is explicit and never reroutes safety refusals, cancellation, permission failures, or operations with an unknown upstream outcome:

imagegen-bridge generate "A red paper fox" \
  --provider codex-responses \
  --fallback codex-app-server:gpt-image-2 \
  --fallback-policy on_unavailable

Docker

git clone https://github.com/Crimsab/imagegen-bridge.git
cd imagegen-bridge
export IMAGEGEN_BRIDGE_BEARER_TOKEN="$(openssl rand -hex 32)"
export IMAGEGEN_BRIDGE_CODEX_HOME="$HOME/.codex"
docker compose up --build -d

The container runs as UID/GID 10001, supports a read-only root filesystem, and keeps OAuth state, SQLite data, artifacts, and reference inputs in separate mounts. Tagged releases publish Linux AMD64 and ARM64 images at ghcr.io/crimsab/imagegen-bridge. Read docs/deployment.md before exposing the API or mounting Codex credentials.

Configuration

Configuration precedence is:

defaults < TOML file < IMAGEGEN_BRIDGE__* environment < --set/--unset

Start with imagegen-bridge setup, or use config.example.toml for a hand-managed deployment. Unknown keys fail validation. config show and config origins report the effective configuration without resolving credential values.

Documentation

• CLI reference
• HTTP API contract
• SDK guide
• Deployment and operations
• Testing matrix
• Release process
• Contributing

Testing

cargo fmt --all --check
cargo clippy --locked --workspace --all-targets -- -D warnings
cargo test --locked --workspace --all-targets

Offline tests use fake Codex processes, mock HTTP services, fixture images, and the production dashboard. Live OAuth tests are individually gated because they can consume image-generation allowance. See docs/testing.md.

Security and upstream status

The default Responses adapter is live-tested against the Codex backend. Its wire protocol is an implementation detail of Codex OAuth and may evolve, so releases keep regression fixtures and controlled live gates. Imagegen Bridge does not disable or bypass upstream safety checks and does not automatically retry an unchanged safety refusal.

Do not commit auth.json, mount an entire home directory into the container, or bind an unauthenticated bridge to a public interface.

License

Licensed under the MIT License.