Multi-provider, multi-tenant image generation on Cloudflare's edge — pick Gemini Lite, Flash, Pro, or OpenAI GPT Image 2 / 2.5 per request.
Six models — Gemini 3.1 Flash Lite (Nano Banana 2 Lite — default), Gemini 3.1 Flash (Nano Banana 2), Gemini 3 Pro Image (Go Bananas! Pro), OpenAI GPT Image 2, and OpenAI GPT Image 2.5 Flare and Sunburst. Tenant-scoped allowlists with per-request `model_id` override.
Generation, editing, Higgsfield video handoff, conversational refinement, character library, products, scenes, reference groups, style presets, search, and analytics — all surfaced through the Model Context Protocol and a parallel REST API.
Save a character (or up to 5) once. Generate consistent scenes across sessions with image-1 identity-locking and base-prompt inheritance.
"Add clouds. Make it brighter. Change to night." The session tracks `last_image_id` so each follow-up edits the previous output without re-uploading anything.
Save product images from URLs, generate consistent marketing materials at scale across unlimited scenes — with per-scene character + product placement.
300+ Cloudflare locations, sub-10ms auth latency, no cold starts. Multi-tenant by design with AES-GCM-encrypted credentials per provider.
Same prompt. Three providers. One model_id.
{
"prompt": "A red fox in a winter forest, golden hour",
"model_id": "gemini-flash-image"
}Each provider implements the same ImageProviderClient contract — circuit breaker, exponential-backoff retry, aspect-ratio enforcement — so swapping model_id is the only change you make. See the OpenAI gpt-image-2 deep-dive for the parameter surface, or Pro Best Practices for Gemini-specific tips.
Three real outputs. Three providers. One API.
Each card is an actual generation — character work routed to Pro for grounded photorealism, product photography to Flash for speed, typography to gpt-image-2 because that's its strength. Same generate_image endpoint, different model_id.
Every architecture diagram on this site was generated by the platform itself.
| Method | Best For | Description |
|---|---|---|
| Claude Code Skills | AI-assisted development | Native plugin integration for Claude Code |
| REST API | Web & mobile apps | Standard HTTP endpoints for any application |
| MCP Protocol | Claude Desktop, Cursor | Full MCP support for AI assistants |
| CLI | Terminal, scripts, CI/CD | gb generate "sunset" for automation |
All tools are exposed via MCP and REST; the CLI supports core workflows and is adding parity over time.
claude mcp add --transport http -s user go-bananas https://mcp.gobananasai.comcodex mcp add go-bananas --url https://mcp.gobananasai.com
codex mcp login go-bananasRun npx -y gobananas-cli auth login, then add to claude_desktop_config.json or .cursor/mcp.json:
{
"mcpServers": {
"go-bananas": {
"command": "npx",
"args": ["-y", "gobananas-mcp"],
"env": {
"GO_BANANAS_SERVER_URL": "https://gobananasai.com",
"GO_BANANAS_MCP_TRANSPORT": "streamable-http"
}
}
}
}Use either direct HTTP OAuth or the STDIO proxy. Logging into gobananas-cli only authenticates the STDIO proxy, not a separate https://mcp.gobananasai.com entry.
Then ask Claude:
"Generate an image of a sunset over mountains with purple clouds"
"Create a character named 'Luna' — silver hair, blue eyes, red cape"
"Generate Luna playing in a meadow"
| Section | Description |
|---|---|
| Guide | Installation, quick start, authentication, core concepts |
| Features | Image generation, editing, characters, products, analytics — and the new OpenAI gpt-image-2 deep-dive |
| Master Guides | Character Consistency, Product Marketing, Workflow Diagrams |
| Console | Web dashboard for managing images and settings |
| MCP Tools | All 53 MCP tools with parameters and examples |
| REST API | Complete REST API reference with curl examples |
| Architecture | System diagrams, data flows, multi-provider routing |
| Self-Hosting | Deploy your own instance on Cloudflare |