bit-graphics

Social-media image generation and style analysis, powered by the same provider stack as the bit-graphics studio app. Being hosted, generated images come back as inline MCP image blocks — Claude Desktop renders them directly in the conversation.

Tools

generate_social_image

Generate social-media-ready images. 2 variants by default, returned fastest-first.

Param	Type	Notes
prompt	string, required	What to depict, including style guidance
format	enum	Preset controlling aspect ratio — see list_formats. Default ig-square
provider	gemini | openai	Nano Banana (default) or gpt-image. Switch on errors — see Failover
count	1–4	Variants; default 2

Formats: ig-square 1:1 · ig-portrait 4:5 · ig-story 9:16 · li-square 1:1 · li-landscape 16:9 · li-story 9:16 · custom-square/portrait/story/landscape/widescreen

analyze_image_style

Send an image (base64), get structured style JSON back: a short label, a long reusable prompt describing the visual language, a hex palette, typography notes, and the design era. Feed the returned prompt into generate_social_image to produce new imagery in the same style.

Param	Type	Notes
image_base64	string, required	Raw base64, no data: prefix
mime_type	enum	image/png (default), jpeg, webp, gif
provider	gemini | openai	Vision backend; Gemini (default) or OpenAI

list_formats

No parameters — returns the format preset table.

Failover

Both tools take provider. If a call fails (429 rate limit, 5xx, quota), the result has isError: true and a JSON body:

{ "error": "...", "provider": "gemini", "status": 429, "retryable": true,
  "retryWithProvider": "openai", "hint": "gemini failed (429). Call this tool again with provider: \"openai\"." }

Clients (or the model driving them) just repeat the call with provider set to retryWithProvider. It is null when the other provider has no key configured on the server.

Notes

• Generation takes 20–90s depending on variant count; clients should allow long tool timeouts.
• Model selection is server-side via env (GEMINI_IMAGE_MODEL, OPENAI_IMAGE_MODEL, GEMINI_TEXT_MODEL, OPENAI_TEXT_MODEL). Both GEMINI_API_KEY and OPENAI_API_KEY must be set for failover to work.
• Defaults: gemini-3.1-flash-image (Gemini) and gpt-image-2.5-sunburst-2026-09-08 (OpenAI). Override either via env without a redeploy of the code.
• Gemini image generation runs on the Interactions API (ai.interactions.create with response_format: {type: "image"}) via @google/genai@2, not the older models.generateContent + responseModalities path (still present in the SDK but deprecated). analyze_image_style's Gemini vision path still uses generateContent, since that's text-out, not image-out, and remains fully supported.
• The local stdio twin of this server lives in the bit-graphics repo (scripts/mcp-server.mjs) — it writes files to disk instead of returning inline images, and its analysis uses the studio app's exact preset vocabulary. If you update its provider SDKs too, note it's on the older models.generateContent image path unless you port this change over.