MCP API — Image Generation

Overview

This section covers generating or uploading images via MCP to use as reference assets in the video pipeline. The core tools are upload_asset and generate_image; the returned assetId can be reused in create_clip and other tools.

upload_asset

Takes an external image URL or base64 data and stores it stably in Autoagens Storage. Free.

Parameters

• url — public image URL
• base64Data — or base64 data (data URL supported)
• mimeType — required when base64Data is raw (e.g. image/png)

Returns

• assetId — stable ID starting with mcpa_...
• publicUrl — Supabase Storage public URL
• mimeType, width, height

Use Cases

• Bring existing brand logos / product photos into the video pipeline
• Persist externally-generated images as stable public URLs (avoid session URL expiry)
• Pass a character reference image to characterReferences

generate_image

Gemini-based AI image generation. Supports text-to-image and image-to-image (with reference images). 100 coins.

Parameters

• prompt (required) — image generation prompt (Korean or English)
• referenceImages — array of public URLs (image-to-image)
• referenceAssetIds — array of pre-uploaded assetIds (recommended for reuse)

Returns

Same structure as upload_asset — assetId, publicUrl, mimeType, width, height

Use Cases

• Generate a character master image → reuse the same assetId across multiple create_clip calls (consistency)
• Generate variations while preserving the reference image's style/composition
• Generate start/end keyframes for I2V mode

upload_asset vs generate_image

Property	upload_asset	generate_image
Purpose	Store existing image	Create new image
Cost	Free	100 coins
Input	URL or base64	Prompt (+ references)
Returns	assetId + publicUrl	assetId + publicUrl

Character Consistency Strategy

To maintain the same person/product across multiple scenes:

1. generate_image → create master character → get assetId_main
2. For each scene, call create_clip:

   {
     "mode": "image-text-to-video",
     "model": "kling-v3-pro",
     "imageAssetId": "assetId_main",
     "characterReferences": [{ "frontalImageAssetId": "assetId_main" }],
     "prompt": "scene-specific motion description"
   }

3. For multi-scene chaining: use previousClipTaskId + referenceVideoUrl to strengthen motion continuity

Note

Gemini image generation supports both Korean and English prompts. Reference images must be fetchable via public URL; expired or session-dependent URLs will fail.