nanobanana-mcp-server

# Tools / Resources / Prompts — Design spec Below is a concise “contract” for each component so any MCP client (Cursor, Claude Desktop, VS Code, etc.) can use the server consistently. ## Tools ### 1) `generate_image` * **Purpose:** Create one or more images with Gemini 2.5 Flash Image (“nano banana”). * **Params** * `prompt` *(str, required)* — Detailed prompt. Include subject, composition, lighting/camera, style; add an **aspect hint** (e.g., “Square image”, “16:9”) in text. ([Google AI for Developers][6]) * `n` *(int, 1–4, default 1)* — Requested image count (model may return fewer). * `negative_prompt` *(str, optional)* — What to avoid. * `system_instruction` *(str, optional)* — Optional system tone/style. * `images_b64` *(list\[str], optional)* — Inline base64 input image(s) to compose / edit / style-transfer against. * `mime_types` *(list\[str], optional)* — MIME type per inline image. * **Returns:** * **Content blocks:** a short text summary + 1..N **image blocks** (PNG). * **Structured JSON:** `{ requested, returned, negative_prompt_applied, used_inline_images, images:[{response_index, image_index, mime_type, synthid_watermark}] }` * **Notes:** * Inline images use `Part.from_bytes`; prefer **Files API** for >20MB or reuse. ([Google AI for Developers][3]) * The model produces interleaved text & images; we extract image `inline_data` parts. ([Google AI for Developers][6]) * All output images carry a **SynthID** watermark. ([Google AI for Developers][5]) ### 2) `edit_image` * **Purpose:** Precise, **conversational** editing (add/remove/mask/change) while preserving original style & lighting. ([Google AI for Developers][6]) * **Params** * `instruction` *(str, required)* — e.g., “Add a knitted wizard hat on the cat; match the soft lighting.” * `base_image_b64` *(str, required)* — Base64 of the image to edit. * `mime_type` *(str, default `image/png`)* — MIME type of base image. * **Returns:** * **Content blocks:** a short text summary + the edited image(s). * **Structured JSON:** `{ returned, synthid_watermark: true }` * **Notes:** * The edit instruction style mirrors official templates for “Adding/removing elements”, “In-painting”, etc. ([Google AI for Developers][6]) ### 3) `upload_file` * **Purpose:** Upload a local file to Gemini **Files API** and get a reusable `uri` (better for large inputs or repeated use). * **Params** * `path` *(str, required)* — Server-accessible path. * `display_name` *(str, optional)* * **Returns:** `{ uri, name, mime_type, size_bytes }` * **Notes:** Inline bytes are limited by a **20MB** request cap; Files API is recommended beyond that or for reuse. ([Google AI for Developers][3]) ## Resources ### `gemini://files/{name}` * **Purpose:** Fetch **Files API** metadata by `name` (e.g., `files/abc123`). * **Returns:** `{ name, uri, mime_type, size_bytes }` * **Rationale:** Gives clients a stable way to introspect file handles they plan to reuse across prompts. ([Google AI for Developers][3]) ### `nano-banana://prompt-templates` * **Purpose:** A machine-readable catalog of prompt templates (the same as the server’s `@mcp.prompt` items). * **Returns:** A JSON dictionary listing the template names, descriptions, and parameter lists. > FastMCP resources are simple functions; returning a dict produces JSON content per the Resources guide. ([FastMCP][7]) ## Prompts (reusable templates) > FastMCP `@mcp.prompt` returns **string messages** that clients can pass straight to tools. ([FastMCP][8]) 1. **`photorealistic_shot(subject, composition, lighting, camera, aspect_hint)`** * Captures photo realism & camera/lighting specifics. Include an **aspect hint** (“Square image”, “16:9”) in text because Gemini infers size/aspect from text cues. ([Google AI for Developers][6]) 2. **`logo_text(brand, text, font_style, style_desc, color_scheme)`** * Aims for **accurate text rendering** in logos; prompt fields mirror the official “text in images” guidance. ([Google AI for Developers][6]) 3. **`product_shot(product, background, lighting_setup, angle, aspect_hint)`** * Studio product mockups with high realism; aligns with recommended product template. ([Google AI for Developers][6]) 4. **`sticker_flat(character, accessory, palette)`** * Flat/kawaii sticker with bold lines and a **white background** (explicit in prompt to enforce cutout-friendly output). ([Google AI for Developers][6]) 5. **`iterative_edit_instruction(what_to_change, how_it_should_blend)`** * Short instruction phrasing for **conversational editing** that preserves original style/lighting. ([Google AI for Developers][6]) 6. **`composition_and_style_transfer(target_subject, style_reference, style_desc)`** * For blending or restyling with multiple inputs; pairs with `images_b64` or Files API URIs. ([Google AI for Developers][6]) ---