Core Concepts

Understanding these core concepts will help you get the most out of Go Bananas!.

Model Context Protocol (MCP)

What is MCP?

MCP (Model Context Protocol) is an open standard for connecting AI assistants to external tools and data sources. It defines how AI models like Claude can:

• Discover available tools and their capabilities
• Execute tool calls with structured parameters
• Receive structured responses

Why MCP?

Before MCP, every AI integration required custom code. MCP provides:

• Standardization: One protocol for all AI-tool integrations
• Discovery: AI can ask "what can you do?" and get a structured answer
• Type Safety: Parameters and responses are schema-validated
• Composability: Combine multiple MCP servers for complex workflows

MCP in Go Bananas!

Go Bananas! exposes 53 MCP tools across categories:

Generation, Editing, Posters, Characters, Products, Style Presets, Reference Groups, Scenes, Models, Analytics

View all 53 MCP tools →

---

Prompting Philosophy

The Golden Rules

Go Bananas! Pro is a "Thinking" model. It doesn't just match keywords; it understands intent, physics, and composition. Stop using tag soups and start acting like a Creative Director.

Golden Rules of Prompting

Rule	Description
Edit, Don't Re-roll	If 80% correct, use continue_editing instead of regenerating
Natural Language	Write full sentences like briefing an artist, not tag soups
Materiality	Describe textures: "brushed steel", "soft velvet", "crumpled paper"
Context	Add "for whom": "for a luxury cookbook" changes everything
Identity Locking	"Keep facial features exactly the same as Image 1"

Good vs Bad Prompts

Bad (Tag Soup):

dog, park, 4k, realistic

Good (Natural Language):

A golden retriever playing fetch in a sunny park, captured in photorealistic detail

Learn Pro Model Best Practices →

Negative Prompting

Tell the model what NOT to include using the negative_prompt parameter.

Common Negative Prompts:

Negative	Why
no date stamp	Pro adds timestamps to corners
no text	Clean images without text
not rustic	Model tends to age things
no watermarks	Prevents generated watermarks

{
  "prompt": "Modern minimalist kitchen",
  "negative_prompt": "not rustic, no wood grain, no farmhouse style"
}

Prompt Evolution

Start simple and iterate with continue_editing:

1. generate_image({ prompt: "A portrait" })
   → Model shows defaults

2. continue_editing({ prompt: "make it dramatic" })
   → Adds contrast, intensity

3. continue_editing({ prompt: "add studio lighting" })
   → Professional photography style

4. continue_editing({ prompt: "avant-garde fashion" })
   → Final creative direction

Edit, Don't Re-roll

If 80% correct, use continue_editing instead of regenerating!

---

Tenants

What is a Tenant?

A tenant is an isolated account within Go Bananas!. Each tenant has:

• Unique ID: acme-corp, creative-studio, etc.
• API Keys: One or more keys for authentication
• Encrypted Credentials: Provider API keys (Gemini and/or OpenAI) stored separately, AES-GCM encrypted in tenant_provider_credentials
• Allowed Models: allowedModels list controls which providers each tenant can route through
• Quotas: Storage limits, rate limits
• Data: Images, characters, presets - all isolated

Multi-Tenant Architecture

Complete data isolation between tenants

Tenant Isolation

• Database: All tables have tenant_id foreign keys
• Storage: R2 keys prefixed with {tenant_id}/
• Encryption: Each tenant's provider keys are individually encrypted; revoking one provider doesn't affect the others
• Quotas: Usage tracked per-tenant

---

Sessions

What is a Session?

A session represents a conversation context for image operations. It tracks:

• Last Image ID: For conversational editing
• Total Images: Count of images in session
• Total Edits: Count of edit operations
• Activity: Last activity timestamp

Session Lifecycle

New → Active → Editing → Inactive states

Why Sessions Matter

Sessions enable conversational editing:

User: Generate a sunset image
Bot: [Creates Image A, session.lastImageId = A]

User: Add some birds
Bot: [Edits A → creates B, session.lastImageId = B]

User: Make it more dramatic
Bot: [Edits B → creates C, session.lastImageId = C]

Without sessions, each edit would require specifying which image to edit.

Session vs Image History

• Session: Temporary working context
• Image History: Permanent record of all generated images

---

Characters

What is a Character?

A character is a saved design that maintains visual consistency across multiple image generations.

Character Components

Component	Purpose
character_name	Unique identifier
base_prompt	Core appearance description
reference_image_ids	Example images for consistency
negative_prompt	What to avoid
system_instruction	Style guidance
tags	Organization

How Character Generation Works

Load character, build prompt, generate with references

Best Practices

1. Detailed base prompt: Include all distinctive features
2. Reference images: Add 2-3 good examples
3. Consistent style: Use similar system instructions
4. Test variations: Generate test scenes to verify consistency

---

Style Presets

What is a Style Preset?

A style preset is a saved template of prompt settings for consistent styling across generations.

Preset Components

Component	Applied How
prompt	Prepended to user prompt
negative_prompt	Added to negative prompt
system_instruction	Sets model behavior
aspect_ratio	Default ratio

Example: "Vintage Film" Preset

{
  "name": "Vintage Film",
  "prompt": "vintage 35mm film photograph, grain texture",
  "negative_prompt": "digital, sharp, modern, oversaturated",
  "system_instruction": "Create images with the aesthetic of 1970s film photography",
  "aspect_ratio": "4:3"
}

When used:

• User prompt: "a coffee shop"
• Final prompt: "vintage 35mm film photograph, grain texture, a coffee shop"

---

Image Lineage

Edit History

Every edited image maintains a link to its parent:

Version branching and lineage tracking

Tracking Lineage

Field	Purpose
parent_image_id	Previous version
edit_depth	How many edits deep
operation_type	generate, edit, character, product

This enables:

• Version history: See how an image evolved
• Branching: Create variations from any point
• Attribution: Track original sources

---

Storage Architecture

Hybrid R2 + D1

R2 for images, D1 for metadata and queries

Why Hybrid?

Storage	Best For
R2	Large binary files, CDN delivery
D1	Queries, relationships, search

Storage Paths

Images are stored with tenant isolation:

{tenant_id}/{date}/generate-{uuid}.png
{tenant_id}/{date}/generate-{uuid}-thumb.jpg

---

Gemini Files API

Temporary Caching

When Gemini generates an image, it returns a gemini_file_id that can be reused for 48 hours:

Fast edits within 48 hours, automatic fallback after expiry

Benefits

• Faster edits: No re-upload needed within 48 hours
• Lower latency: Skip R2 round-trip
• Automatic fallback: Gracefully handles expired files

---

Next Steps

Now that you understand the core concepts:

• Explore all features →
• Learn MCP tools in detail →
• View the architecture →