MCP Setup

What this is

Claude can't render images or videos on its own. Claude Imagine's MCP server closes that gap: connect it once, then ask Claude for an image or video in plain English and it generates one using your credit balance — no leaving the chat.

The server speaks the Model Context Protocol over Streamable HTTP at:

https://claudeimagine.com/api/mcp

Authentication is OAuth 2.1 (PKCE + dynamic client registration). There is no API key to paste. When you add the server, your Claude client discovers the auth endpoints automatically and opens a browser window for you to log in and consent. After that, it just works.

New accounts start with 9 credits. The first authenticated MCP connection automatically unlocks a one-time 6-credit bonus, valid for 14 days, so you can earn up to 15 free credits in total.

Setup for Claude Code (CLI)

Run one command:

claude mcp add --transport http claude-imagine https://claudeimagine.com/api/mcp

Claude Code opens a browser to log in and consent. Once you approve, the ten tools below are available in your sessions.

Setup for claude.ai and Claude Desktop

Add it as a custom connector by URL. The feature is in beta and available on Free, Pro, Max, Team, and Enterprise plans (Free is limited to one custom connector).

Pro / Max (individual):

1. Go to Customize → Connectors.
2. Click +, then Add custom connector.
3. Paste https://claudeimagine.com/api/mcp as the remote MCP server URL.
4. Click Add. A browser window opens for OAuth login and consent.

Team / Enterprise: an owner first adds it under Organization settings → Connectors (Add → hover Custom → Web → paste the URL → Add), then each member goes to Customize → Connectors, finds the connector marked "Custom", and clicks Connect to authenticate.

Once connected, enable it per conversation with the + button in the chat, then Connectors.

Menu wording can shift as Claude's UI evolves. The constant is: Connectors → add a custom connector → paste the URL above.

Tools

Once connected, Claude has ten tools:

generate_image

Generate an image from a text prompt.

• prompt (string, required) — what to draw.
• model (optional) — one of nano-banana-2 (default, best quality/price), gpt-image-2, seedream-4.5, flux-2-pro, z-image (fastest & cheapest).
• aspect_ratio (optional) — one of 1:1, 4:3, 3:4, 16:9, 9:16. When omitted, the selected model's own default is used.
• style (optional) — a named look from list_styles, for example thumbnail-bold. The prompt then describes only the subject; the style supplies the framing and its category's aspect ratio unless aspect_ratio is given.
• max_credits (optional) — a budget cap in credits. If the exact cost is above it, nothing is started and the reply lists what fits.

Returns the image inline (a JPEG preview) plus the URL of the full-resolution file. In claude.ai, Claude Desktop and the Claude mobile apps the image renders as an inline card (MCP Apps); clients that do not support MCP Apps, such as Claude Code, receive the image block and the link. Credits are deducted when the task starts and restored if it fails, is canceled, or produces no output.

The card carries Open full size, Download and Edit buttons where the client supports them. Download asks the client to save the full-resolution file; Edit takes a one-line change ("make the background white") and posts it into the chat, so Claude runs edit_image with the conversation's context rather than the card guessing. A result made with a style also offers Try another style, which asks Claude for the rest of that category.

edit_image

Change an existing image with a text instruction. The result is a new image; the original is untouched.

• prompt (string, required) — what to change, and what to keep. Saying what must stay the same gives better results than naming only the change.
• image_urls (array, required) — one public image URL, normally the full-resolution URL returned by an earlier generation. One image per call in this version.
• model (optional) — nano-banana-2 (default) or gpt-image-2. The other image models cannot edit through MCP yet; use the web editor for those.
• aspect_ratio (optional) — same five values. When omitted, the framing of the source image is kept.
• max_credits (optional) — a budget cap in credits. If the exact cost is above it, nothing is started and the reply lists what fits.

Credits follow the same rule as generate_image: deducted when the task starts, restored if it fails, is canceled, or produces no output. If the image is only on your device, upload it in the web editor instead; the connector needs a URL Claude Imagine can reach.

list_styles

Named looks for generate_image, four per category, free to call:

• category (optional) — only that category. Omit it for all six.

Each style is a prompt recipe; your prompt is the subject dropped into it. In claude.ai, Claude Desktop and the mobile apps the styles appear as a picker card: click a tile and Claude generates with that style. Elsewhere Claude lists them and you name one. Words that must appear in the image go in the prompt in double quotes, for example a tiny FPV drone, title "Tiny drone, big power".

generate_video

Generate a video with one of six server-priced models.

• prompt (string, required) — the scene, motion, camera movement, lighting, and style.
• model (optional) — h3-max-turbo (default), grok-imagine, seedance-1.5-pro, seedance-2.0-mini, veo-3.1-fast, or kling-2.5-turbo.
• duration (optional) — H3 Max Turbo accepts 5–15 seconds; Grok accepts 6–30; Seedance 1.5 Pro accepts 4–12; Seedance 2.0 Mini accepts 4–15; Veo 3.1 Fast accepts 4, 6, or 8; Kling accepts 5 or 10.
• resolution (optional) — H3 Max Turbo supports 480p or 768p; Grok and Seedance 1.5 support 480p, 720p, or 1080p; Seedance Mini supports 480p or 720p; Veo Fast currently uses 720p; Kling has fixed Pro quality.
• image_url (optional) — one public image URL to animate. Omit it for text-to-video.
• aspect_ratio (optional) — one of 16:9, 9:16, 1:1. Defaults to 16:9; Veo accepts 16:9 or 9:16, and Kling or H3 Max Turbo image-to-video follows the source image.
• max_credits (optional) — a budget cap in credits. If the exact cost is above it, nothing is started and the reply lists what fits.

The tool calculates the price from the selected model, duration, and quality. When the render finishes, claude.ai, Claude Desktop and the Claude mobile apps play the clip inline (MCP Apps); other clients receive the MP4 URL. Failed generations are refunded automatically. All combinations and prices are independently validated server-side.

check_generation

Returns the result of a generation that was still running when its tool call ended. Claude calls this itself: a generate_image, edit_image or generate_video reply that says the task is still running quotes a task id and asks Claude to check it, so you never have to leave the conversation or look anything up.

• task_id (string, required) — the id quoted in the still-running reply.

It waits for the task, then answers exactly as the original tool would have. It consumes no credits. It also returns a finished generation again: every generation reply quotes its task id, so "show me that image again" is a free call rather than a new generation.

quote_generation

The exact credit cost of a generate_image, edit_image or generate_video call before it is made, together with your balance and whether your plan allows the model. Claude uses it whenever price or budget comes up, so the number it quotes is the number you are charged.

• tool (string, required) — which call is being priced.
• model, duration, resolution, with_image (optional) — the same values the call would use.

Free; it reserves nothing and starts nothing.

list_generations

Your own recent image and video generations on this account, newest first, including the ones you made on the website. Each entry carries its task id and file URL, so in a new conversation Claude can find "the product image from yesterday" and show it again, edit it or animate it.

• query (optional) — text to look for in the prompt.
• type, status, limit (optional) — narrow the list; at most 20 entries.

Free; it returns records only.

list_models

Lists the available image and video models and their live credit costs for your current plan tier. Takes no arguments.

get_credits

Returns your remaining credit balance. Takes no arguments.

plan_product_ad

Plans a whole product-photo workflow: two image directions from one original photo, your choice between them, then a short video. It returns an itemised quote at your account's prices, the prompts for each step and an output checklist. Free, and it starts nothing. The worked example is at the end of this page.

Choose a video model in plain English

You do not need to memorize a model ID or write tool arguments. Say the model family you know and Claude Imagine maps it to the current supported model:

If you do not name any model, generate_video uses H3 Max Turbo by default. Every account's first H3 Max Turbo clip at 5 seconds and 480p is free; quote_generation returns 0 credits for it while it is unclaimed. Exact IDs remain available for advanced workflows:

Check models and cost before spending credits

Ask Claude to call the read-only tool first:

Check the available video models and current credit costs, but do not generate anything yet.

Claude calls the read-only list_models tool. It does not create a task or consume credits.

Text to video with Veo 3.1 Fast

Use Veo to make an 8-second landscape video: A slow cinematic push through a rain-soaked neon alley, reflections moving across the pavement, with natural camera motion.

Text to video with Seedance 2.0 Mini

Use Seedance to make a 10-second 720p vertical video: A clean product reveal with a slow orbiting camera, soft studio light, and restrained motion.

Generate an image with a specific model

Image generation uses the same pattern:

Use Nano Banana 2 to make a square image: A minimal studio product photograph on warm gray paper, soft window light, realistic texture.

Make a YouTube thumbnail in a named style

Show me the YouTube thumbnail styles, then make one for my video about building a tiny FPV drone

Claude calls list_styles with the youtube-thumbnail category, shows the four looks, and after you pick one calls generate_image with that style and a prompt that describes the subject. The 16:9 framing comes from the category.

Edit an image you just generated

Generate a 16:9 photo of a lighthouse at dawn, then add a flock of birds to the sky

Claude calls generate_image, then passes the returned full-resolution URL to edit_image. Each step costs its own credits, and the edit keeps the 16:9 framing unless you ask for a different one.

Animate an image

Add image_url to the same request:

Use Veo to animate https://example.com/my-image.jpg as a 6-second vertical video. Preserve the subject and composition while the camera slowly pushes in and the background moves with subtle parallax.

image_url must be a publicly reachable HTTP or HTTPS URL. A local file path or a chat attachment is not automatically a public URL; if Claude cannot provide a reachable URL, use the web generator to upload the image instead.

Credits

MCP and the web generators share one balance — the same credits work everywhere, including the 6-credit MCP activation bonus. The bonus is granted once per account and expires 14 days after activation. You top up on the same pricing page; image and video costs depend on the selected model. Call list_models to see current costs for your account.

Continue with Claude Imagine

• Connect the MCP server in Claude
• Create videos with Claude, Veo, Seedance, or Kling
• Follow the Claude image MCP guide
• Compare subscriptions and credit packs

Troubleshooting

• The OAuth popup didn't appear. Re-run the add command (Claude Code) or remove and re-add the connector (claude.ai / Desktop). The login window only opens on the first connection.
• "Not enough credits." Check your balance with get_credits and top up on the pricing page.
• Costs look different than expected. Credit costs shown by list_models are for your current plan tier — they change if your plan changes.

A note on independence

Claude Imagine is an independent tool that works alongside Claude. Claude is a trademark of Anthropic, PBC. We are not affiliated with, endorsed, or sponsored by Anthropic.

Product photo → two directions → short video

Open the Product Ad workshop to upload a product photo and describe the packaging details to preserve. Copy to Claude includes the reachable photo URL, product brief and total credit limit. A chat attachment on its own is not a public URL. This starts a separate workflow in Claude; it does not transfer a partially completed browser draft.

Ask Claude to call plan_product_ad first:

{
  "product": "LUMA Daily Serum",
  "video_model": "seedance-1.5-pro",
  "image_url": "https://your-public-image-url.example/product.jpg",
  "preserve": "Exact label, bottle color and dropper shape",
  "max_credits": 12
}

The tool returns a live itemized quote, two image-edit prompts, a video prompt and an output checklist. It does not generate or spend credits. Set video_model to any supported image-to-video model and duration to its supported clip length: H3 Max Turbo 5–15 seconds, Grok Imagine 6–30, Seedance 1.5 Pro 4–12, Seedance 2.0 Mini 4–15, Veo 3.1 Fast 4/6/8, or Kling 2.5 Turbo 5/10. The plan uses each model's default quality (480p, 720p for Veo, or fixed Pro for Kling). If the model is omitted, it selects the account's free H3 preset when eligible, otherwise the cheapest preset supporting the requested length. An explicitly chosen model stays selected after an offer is used; the new paid price must fit the approved budget. The free H3 offer only covers 5 seconds at 480p. The website's Copy to Claude includes the selected model and duration.

After approving the quote, generate the two image directions sequentially, each from the original product photo. Choose one completed result before making the video. Claude should re-quote before every step, keep task IDs, track costs including retries, and set each generation's max_credits to the smaller of the approved step cost and the remaining overall budget. The planning tool supplies this guidance; it does not create a shared server budget across multiple tool calls. Each generation's own cap is checked against its actual charge before credits are reserved, including when a free-clip offer is no longer available.

Finish by inspecting every label, comparing product shape and color with the original, and watching for warping or changing text through the entire clip. Prompts cannot guarantee fidelity. Keep the original result URLs; list_generations and check_generation let you continue later without generating duplicates.