CrossGen

CrossGen 0.3.4 is released: it adds GPT Image 2.5 launch support, a media-aware foundation, and a confirmed AppLink import flow on top of the reliable 0.3.3 image workspace. v0.3.5 is in development and consolidates every feature and fix merged after that release: the GPT Image 2.5 Sketch input workspace and Input Studio integration, the strict model-discovery evidence contract (GPT Image 2 and 2.5 separated by exact provider model ID), the canonical Gemini image-model contract, capability-gated General image-to-image support, and the planned video technology preview. The visual app, CLI, and MCP keep sharing the same queue, provider diagnostics, reference-image handling, History, Gallery, and media metadata contracts, and the bundled crossgen-artist skill stays on the same path. v0.3.5 is not released yet; the download link still points at v0.3.4.

The desktop app, CLI, and MCP server share the same API profiles, durable generation queue, History, and Gallery. The bundled skill stays aligned with that contract. Install CrossGen once; installed CLI and MCP use require no separate Node.js, npm, pnpm, global package, or local HTTP service.

Agent Quickstart

Install and open the latest desktop release, then add an API profile in API access. If you are using Codex or another agent, install the bundled CrossGen Artist Skill first; it keeps model discovery, queue jobs, and export behavior aligned with the app. The packaged crossgen launcher is ready for scripts and can start the local MCP stdio server with --mcp.

crossgen doctor --agent --json
crossgen models list --json
crossgen generate --prompt "A precise isometric app icon" --model gpt-image-2.5-sunburst --quality xhigh --yes --wait --json

Generate least-privilege MCP configuration for your client:

crossgen mcp config --client codex --mode readonly --json
crossgen mcp config --client claude-code --mode generate --json
crossgen mcp config --client cursor --mode generate --json

Start with readonly, then enable write or generate only when the agent needs it. Paid generation, asset export, destructive actions, queue-control changes, and local path disclosure require explicit permission or confirmation. See the CLI and MCP guide for commands, tool coverage, modes, and installed executable paths.

Codex Plugin

CrossGen ships a Codex plugin at plugins/crossgen and registers it in the repository marketplace at .agents/plugins/marketplace.json. The plugin bundles the crossgen-artist skill together with a stdio MCP server that launches the installed CrossGen app, so Codex can discover models, submit queued jobs, and inspect Gallery assets through the same local runtime as the desktop app. CROSSGEN_MCP_MODE defaults to readonly.

CrossGen Artist Skill

CrossGen ships a first-class, media-aware agent workflow skill at skills/crossgen-artist. It is the recommended entry point for Codex and other compatible agents that need to discover model capabilities, choose generation/edit/inpaint operations, submit idempotent queue jobs, poll completion, inspect image/GIF/video metadata, and export results without exposing API keys or local paths by accident. The released v0.3.4 still does not expose real video generation or video editing; the planned v0.3.5 video technology preview does not promise a full video product either.

Install it for Codex from a CrossGen checkout:

mkdir -p "$HOME/.codex/skills"
ln -sfn "$(pwd)/skills/crossgen-artist" "$HOME/.codex/skills/crossgen-artist"

Then invoke it as $crossgen-artist. The skill is versioned with CrossGen so its instructions stay aligned with the bundled CLI/MCP contract, including GPT Image 2.5 model selection, Responses multi-turn edits, mask preflight, and media-aware Gallery reads. When CrossGen is installed as a packaged app, use the skill files from the matching source/release archive; a future installer will make this one click from Agent access.

The desktop Agent access panel remains the source of truth for MCP setup. Copy a client-specific snippet there, choose readonly for discovery, and switch to write or generate only for workflows that need those permissions. CrossGen does not silently edit Codex configuration or enable paid generation.

AppLink API Configuration

CrossGen 0.3.4 ships a crossgen:// AppLink for API aggregation platforms. An aggregator can open a provider import link and hand CrossGen its API name, Base URL, API Key, and optional model in one step. CrossGen shows a confirmation dialog before saving anything; after receiving the link, it stores the key with the existing local key protection and does not write it to logs or render it in full.

Canonical form:

crossgen://provider/import?name=AIHub&kind=custom&base_url=https%3A%2F%2Fapi.example.com%2Fv1&api_key=sk-...&model=flux-pro

Accepted aliases include apiKey/api_key, baseURL/base_url, defaultModel/default_model, and provider/kind. Supported launch targets are gpt-image-2, gpt-image-2.5, nano-banana-3, and general; an AppLink model hint only seeds the initial config. After import, the latest successful discovery for that API Key is authoritative, and a launch is enabled only when the exact provider model ID is returned. nano-banana-3 is a CrossGen launch/workflow alias, not a Gemini wire model id: pass a discovered provider id such as gemini-3.1-flash-image, gemini-3.1-flash-lite-image, or gemini-3-pro-image when selecting a specific Gemini model. Packaged builds register the crossgen scheme on macOS, Windows, and Linux.

Using CrossGen in a real desktop, CLI, or agent workflow? Star the repository so other local-first image-tool users can find it, and share the workflow in Discord.

It is built for real image-generation work, not just one-off prompting. Designers, comic and storyboarding teams, UI makers, operators, product teams, and AI image hobbyists often need the same loop:

1. connect an image model with an API key,
2. generate a batch of ideas,
3. keep useful outputs,
4. crop, annotate, pick colors, or make quick edits,
5. reuse the result as a reference image,
6. generate again.

CrossGen keeps that whole loop inside one app. No repeated file hunting, no scattered browser downloads, no separate folder cleanup before the next image-to-image attempt.

GPT Image 2 and 2.5

CrossGen v0.3.5 (in development) keeps GPT Image 2 and GPT Image 2.5 as separate launch families. A familiar model name is never treated as proof that the current API Key can use it:

• gpt-image-2 is the GPT Image 2 launch target.
• gpt-image-2.5 is the CrossGen launch/AppLink compatibility alias. It is not a provider capability claim and is never injected as discovery evidence.
• gpt-image-2.5-sunburst and gpt-image-2.5-flare are concrete GPT Image 2.5 provider targets; valid dated variants are accepted only when returned by the provider.
• gpt-image-2.5-sunburst is intended for precise edits, structure preservation, and high-fidelity reference work.
• gpt-image-2.5-flare is intended for fast, high-quality everyday generation.
• Valid dated base, Sunburst, and Flare snapshots are accepted when the provider exposes them.

The latest successful model discovery for the selected API Key is the availability source of truth. CrossGen enables a GPT Image 2 or 2.5 launch only after the exact provider model ID is confirmed by a lightweight image route. Provider-listed rows may use a non-empty 2xx image response as confirmation even when the optional metadata route is unavailable. If a gateway omits image deployments from /models, CrossGen may try a small set of product-owned route candidates; those candidates are launchable only when the successful payload echoes the exact model ID (or exact metadata identity is paired with a reachable validation route), and the UI labels them as route-confirmed rather than provider-listed. Empty success envelopes such as { data: [] }, { output: [] }, or { choices: [] } only prove route reachability and cannot enable a speculative candidate. A validation-only 400/422 is accepted only when the metadata endpoint also echoed the same ID, so discovery does not need to start a paid image generation. A different echoed ID is rejected, and a response without an exact ID echo remains inconclusive for route candidates. A row that is merely listed remains visible but disabled; model_not_found from the metadata endpoint alone is inconclusive, while model_not_found from every exact image route is rejected. A missing metadata route or transport failure is also inconclusive. Explicit image_generation=false, text-only, audio-only, or video-only declarations also veto a launch. If discovery returns only gpt-image-2, CrossGen shows only GPT Image 2; it does not silently add GPT Image 2.5. If it returns only a 2.5 ID, CrossGen shows only GPT Image 2.5. Replacing the API Key clears the previous discovery result before the new probe completes. A failed or stale probe also clears the active model and route until a later successful probe confirms a new exact model ID.

For focused launches, CrossGen derives the family and variant label from the exact provider model ID. A gateway-supplied displayName is advisory, so it cannot relabel gpt-image-2 as GPT Image 2.5 (or the reverse). The same rule applies to concrete 2.5 provider IDs and valid dated snapshots: when the provider returns one, CrossGen keeps that exact ID for routing and shows it alongside the friendly launch label. The bare gpt-image-2.5 value remains a compatibility alias for imported links and older drafts; it does not enable the 2.5 launch. For OpenAI-compatible metadata, name is treated the same way: it may be a product label, so only id, model, model_id, or modelId can confirm an exact GPT Image deployment. Native Gemini metadata still uses name: models/<id> as its canonical resource ID. When discovery promotes the active model, the desktop editor parameters are synchronized to that confirmed model before the next run. GPT Image 2.5 base, Sunburst, and Flare probes are aggregated independently; a confirmed variant never enables a different variant.

Cross-protocol discovery gives the configured transport precedence. When the primary protocol returns at least one runnable image model, its catalogue alone drives the model picker; results from the alternate protocol are not merged. The alternate protocol is used only when the primary response has no runnable image model, and the inferred provider kind is recorded for routing. This keeps protocol-incompatible or error-shaped gateway responses from enabling the wrong focused launch.

OpenAI route evidence is bound to the exact provider model ID that was probed. Switching between gpt-image-2, gpt-image-2.5, or a dated 2.5 variant clears the old route evidence until the newly selected model is probed. The read-only pnpm probe:real-aihub-models command checks GPT Image 2 and GPT Image 2.5 family IDs independently. It also resolves the product launch alias nano-banana-3 to the actual Gemini provider IDs gemini-3.1-flash-image, gemini-3.1-flash-lite-image, and gemini-3-pro-image; it never treats the alias itself as a required Gemini wire model. A gateway that lists a model but returns an explicit model_not_found for /models/{id} is reported as inconclusive until an exact image route confirms or rejects the same provider ID. Only when every probed image route explicitly rejects that exact ID is the result rejected; a generic missing metadata endpoint or transport failure remains inconclusive. Neither state is presented as a launchable model.

The probe JSON keeps the evidence explicit: family is derived from the exact provider ID, exactIdMatch means that ID appeared in /models, metadataStatus describes the per-model endpoint, and only availability: "confirmed" with eligibleForLaunch: true is launchable. Metadata-only responses, ordinary HTTP failures, missing metadata routes, and metadata-only model_not_found responses are inconclusive; metadata that echoes a different model ID is rejected, and an explicit model_not_found from every exact image route is rejected. Only confirmed rows create an enabled model option. The legacy availableImageModels field is retained for readers of older evidence and means “image-looking IDs listed by /models”, not confirmed API-key access. For new consumers, use confirmedImageModels or launchableImageModels for the exact provider IDs that passed the read-only confirmation, and launchableTargetModels for CrossGen launch aliases. The provider-id arrays never contain the nano-banana-3 workflow alias; a Gemini request must use the confirmed gemini-* ID. confirmedTargetModels remains as a compatibility alias for launchableTargetModels. For a launch represented by several provider IDs, matchedProviderModelIds shows the exact IDs returned by /models, and providerProbes preserves each metadata result. This command is read-only and does not replace the desktop runtime's exact-ID and route checks; it intentionally reports metadata uncertainty conservatively.

The desktop, CLI, and MCP surfaces expose the same controls: auto, low, medium, high, xhigh, and max quality; custom 16-multiple dimensions up to the documented 4K envelope; transparent/opaque/automatic background; PNG/JPEG/WebP output and JPEG/WebP compression; n up to 10; and streaming partial images from 0 to 3. Edits support up to 16 reference images, input_fidelity, and mask/inpainting with source/mask format and dimension checks, alpha validation, and a 50 MB mask limit.

An optional user safety identifier is available in the desktop advanced settings, CLI (--user), and MCP (user). CrossGen sends it as user through the Images API and as safety_identifier through Responses. It must be an opaque stable identifier; CrossGen never derives it from API keys or personal account data.

GPT Image 2.5 can use the Image API or the Responses API image-generation tool. Responses requests use a configurable mainline model (default gpt-6-astra) and support action, previous_response_id, revised-prompt metadata, and multi-turn editing. Responses streaming sends partial_images only when partial previews are requested (1 to 3); a value of 0 omits the optional field. GPT Image 2.5 never uses the legacy Chat Completions image route. In automatic route mode, CrossGen uses the Images API for ordinary and batch requests, and uses Responses only when conversation-specific controls are present. Route probes are diagnostic only and do not silently promote a normal 2.5 request to the conversational path. Explicit Responses requests stay on Responses if they fail, preserving conversation semantics instead of retrying through a different billable route.

For privacy and lifecycle simplicity, v0.3.4 sends local Responses inputs as base64 data URLs and does not upload or persist OpenAI Files API IDs. See the GPT Image 2.5 support research note for the full capability matrix and official references.

For agent-driven work, the loop becomes equally direct:

1. inspect configured providers and verified model capabilities,
2. submit text-to-image or reference-image work,
3. track, cancel, or retry the queue-backed job,
4. inspect the resulting Gallery asset,
5. export it into the agent's current project.

Gemini Image Models and Sketch (0.3.5 In Development)

CrossGen keeps the product launch name Nano Banana 3 while preserving the actual Gemini model id used by the provider. The current focused Gemini image ids are:

• gemini-3.1-flash-image — the default Nano Banana 3 launch target;
• gemini-3.1-flash-lite-image — a distinct discovered Gemini image model;
• gemini-3-pro-image — a distinct discovered Gemini image model.

The desktop launch menu, History, CLI, and MCP retain the real provider id for selection and traceability. For OpenAI-compatible gateways, that returned id is the only family-classification source of truth; display_name is advisory. A row such as id: gpt-image-2.5, display_name: GPT Image 2 remains a disabled compatibility row and does not prove GPT Image 2.5 support. Explicit text-only, audio-only, video-only, or image_generation=false metadata can veto an image-looking id. Older drafts and AppLinks containing nano-banana-3 are migrated to gemini-3.1-flash-image; CrossGen never sends the alias as the Gemini request model. The current API key's model-discovery response is authoritative: a model that is not discovered, or whose capability metadata does not confirm image editing and reference images, remains disabled for Sketch and is blocked before a paid request.

Sketch is an image-to-image input workflow inside the existing desktop workspace, not a third top-level mode. The user creates or reopens a Sketch in the reference-image area, draws it in Input Studio, optionally adds a view-only reference underlay, and saves it as the first input image. CrossGen records workflow: "sketch" and Sketch provenance in History/CLI/MCP metadata, while the provider request continues to use the validated image-edit route. CrossGen does not invent or send a provider-native scratch field, and Sketch cannot be combined with a mask.

The code and deterministic/mocked contracts for GPT Image 2.5 and the Gemini focused launches are present in v0.3.5. The final release gate still requires a real AIHub GPT Image 2.5/Nano Banana 3 quality, latency, failure-recovery, and privacy matrix; the presence of Gemini provider IDs confirms the Nano Banana launch mapping but does not replace real workflow acceptance. Mock or compatibility-model responses do not replace that gate.

Visual Tour

Core Workflows

1. One API Key Hub

CrossGen keeps API access simple:

• save multiple API keys and Base URLs,
• switch between OpenAI, Gemini, and OpenAI-compatible providers,
• run model discovery from the app,
• detect available image models,
• automatically probe compatible generation routes,
• keep the active API profile visible without crowding the workspace.

For aggregation platforms, route compatibility matters. CrossGen can prefer the route that actually works for the configured provider, including chat-style image generation paths used by compatible gateways.

2. Gallery And History That Actually Help

Generated images are not throwaway files. CrossGen treats them as reusable working material:

• History records every generation result with prompt, model, duration, and output actions.
• Gallery stores selected images as reusable assets.
• Gallery folders map to local files so assets remain manageable outside the app.
• Tags, folders, search, sorting, and compact/collapsed views keep large libraries usable.
• History and Gallery images can be clicked for preview/editing or dragged into image-to-image references.
• Right-click actions include local path copy for workflows that still need direct file access.

The goal is simple: the image you generated ten minutes ago should be easy to find, edit, and reuse.

3. Generate, Edit, Reuse, Generate Again

CrossGen includes a stronger image preview and editing area:

• crop results and save the selected region as a new image,
• draw quick annotations,
• add text boxes,
• pick colors from the image,
• save edited results into Gallery,
• use Gallery/history items as reference images,
• continue the next image-to-image round without leaving the app.

This makes CrossGen useful for iterative visual work: generate a base image, crop a detail, annotate or adjust it, save it, then feed it back into the next prompt.

Other Highlights

• GPT Image 2 / 2.5 and Gemini image workflows: focused launch entries for GPT Image 2.5 Sunburst/Flare, legacy GPT Image 2, and Nano Banana/Gemini image models.
• Agent-ready CLI/MCP runtime: local agents can discover providers and models, submit queue-backed image jobs, inspect status, and export Gallery assets.
• Aggregation-provider compatibility: release gates include real-provider validation through OpenAI-compatible aggregation endpoints and Gemini-compatible image models.
• Durable generation queue: generation and edit work runs through a bounded local queue with status tracking, retry, cancel, and safe defaults.
• Prompt templates: save reusable prompt structures and apply them quickly.
• Prompt chips: insert Gallery assets, color values, and templates into prompts.
• Image-to-image reference handling: drag local files, Gallery assets, or History outputs into reference slots.
• Dark mode: built for longer visual review sessions.
• Local-first storage: history, outputs, templates, and Gallery assets are stored locally.
• Open source: released under the MIT License.

Agent Runtime

CrossGen exposes the local image runtime through structured JSON CLI commands and an MCP stdio server. The same queue, diagnostics, and Gallery rules protect desktop, CLI, and MCP workflows:

• crossgen doctor --agent --json reports the app path, data directory, provider readiness, queue configuration, and MCP launch hints.
• crossgen mcp config --client codex|claude-code|cursor --mode readonly|write|generate --json prints client-ready MCP configuration.
• crossgen generate ... --yes --wait --json and MCP crossgen_generate_image submit work through the durable queue.
• crossgen asset export <asset-id> --to <path> --yes --json copies a managed image into a project without moving the Gallery source.

CLI and MCP are separate entry points. Packaged MCP configuration uses the bundled CLI launcher with --mcp; on macOS, Windows and Linux it reuses one per-user worker for parallel agent sessions. The launcher does not require Node.js, npm, pnpm, or a global package. CLI and MCP default to read-only behavior. Write and generation modes are explicit, paid generation requires confirmation, and local path disclosure is opt-in.

Generate from a local agent or terminal

After configuring an API profile in the desktop app, use the packaged crossgen launcher:

crossgen doctor --agent --json
crossgen models list --json
crossgen generate --prompt "A precise isometric app icon" --model gpt-image-2.5-sunburst --quality xhigh --yes --wait --json
crossgen job status <job-id> --json
crossgen asset export <asset-id> --to ./assets/app-icon.png --yes --json

Image editing uses the same queue and can accept a local reference image:

crossgen edit \
  --prompt "Keep the composition and change the background to white" \
  --input ./reference.png --yes --wait --json

All machine-facing responses are JSON. Read-only inspection omits API keys and absolute asset paths; paid generation, destructive actions, queue-control changes, exports, and path disclosure require explicit confirmation.

Connect Codex, Claude Code, Cursor, or another MCP host

CrossGen can generate client-ready MCP configuration:

crossgen mcp config --client codex --mode readonly --json
crossgen mcp config --client claude-code --mode generate --json
crossgen mcp config --client cursor --mode generate --json

MCP uses the bundled launcher with --mcp; enabling a global crossgen shell command is not required. Choose the smallest permission mode that fits the workflow:

The packaged macOS/Linux launcher shares one local MCP worker across parallel agent sessions. The runtime reclaims an otherwise idle session after 15 minutes and exits when its host process disappears. Set CROSSGEN_MCP_IDLE_TIMEOUT_MS=0 to disable this guard.

Typical agent workflows include generating UI assets during coding, producing several visual directions and polling them as durable jobs, editing an existing reference image, and exporting selected Gallery results directly into a repository.

See docs/cli-mcp.md for command examples and docs/KNOWN_LIMITATIONS.md for current agent/runtime limits.

Download And Use

CrossGen is distributed as a desktop release package. Download the latest installer from the GitHub Releases page, install it, open the app, add your API key, and start generating.

Basic setup:

1. Open API access.
2. Add an API key and Base URL.
3. Run model discovery.
4. Launch GPT Image 2.5, GPT Image 2, Nano Banana/Gemini, or a compatible model.
5. Generate, edit, save useful images to Gallery, and reuse them as references.

The macOS arm64 release is Developer ID signed and Apple notarized. If Gatekeeper blocks a local unsigned build, right-click the app and choose Open, or clear the quarantine attribute:

xattr -dr com.apple.quarantine /Applications/CrossGen.app

If Windows SmartScreen appears, choose More info and then Run anyway.

Brand

CrossGen means a cross-model, cross-step generation workspace: one local runtime for API access, generation, editing, Gallery management, repeated image-to-image iteration, and AI-agent workflows.

The product promise is deliberately practical:

configure once, generate from the app or an agent, keep useful images organized, and reuse or export every result.

CrossGen is maintained by Nowo and Corgnitor. Nowo focuses on AI-native product design and applied workflows. Corgnitor focuses on AI engineering and productization.

Join the CrossGen community on Discord for feedback, release discussion, and workflow ideas.

Community and Contributions

• Ask usage questions, share workflows, and discuss early ideas in GitHub Discussions.
• Report reproducible problems or scoped requests with the issue templates.
• Read CONTRIBUTING.md before submitting code or documentation.
• Report vulnerabilities privately according to SECURITY.md.

New contributors can start with good first issue or help wanted tasks.

Technical Notes

CrossGen is an Electron + React + Tailwind desktop app. The app focuses on local-first workflows and supports OpenAI, Gemini, and compatible image providers.

Useful commands:

pnpm install
pnpm dev:electron
pnpm build

Validation:

pnpm verify:mock-api
pnpm verify:mock-gemini-api
pnpm verify:mock-model-discovery
pnpm verify:release-evidence

Packaging:

pnpm package:dir
pnpm package:mac
pnpm package:win
pnpm verify:release:mac
pnpm verify:release:windows
pnpm verify:release:linux

Release evidence is tracked in docs/release/evidence.json. Mock verifiers do not spend real API credits. Real-provider gates require explicit cost approval and local environment variables.

License

CrossGen is released under the MIT License.