Visualizee MCP server

Visualizee ships a hosted MCP (Model Context Protocol) server, so any MCP-capable assistant, including Claude, Claude Code, and Cursor, can render with your Visualizee account. Your assistant gets seven tools: it can generate images (with an inline preview right in the chat), video clips and textured 3D models, collect a render that is still processing, upload your local photos as references, look up the available models, and check your balance and plan limits. Everything runs against your own account, and every render is filed into your Visualizee workspace.

The server lives at a single URL:

https://visualizee.app/api/mcp

It speaks Streamable HTTP and authorizes with OAuth 2.1: the first time a client connects, you sign in at visualizee.app and approve access. There are no API keys.

Connect from Claude Code

Add the server from your terminal:

claude mcp add --transport http visualizee https://visualizee.app/api/mcp

Then, inside a Claude Code session, run /mcp and pick visualizee to authenticate. Your browser opens visualizee.app; sign in and approve access. From that point the tools are available in every session.

If you prefer project-level config, add this to .mcp.json in your project root:

{
  "mcpServers": {
    "visualizee": {
      "type": "http",
      "url": "https://visualizee.app/api/mcp"
    }
  }
}

Connect from claude.ai and Claude Desktop

Connect from Lovable

Connect from Cursor

Add the server to .cursor/mcp.json (per project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "visualizee": {
      "url": "https://visualizee.app/api/mcp"
    }
  }
}

Cursor will prompt you to authenticate on first use. Any other MCP client that supports Streamable HTTP and OAuth works the same way: point it at the URL and sign in when asked.

The tools

Tool	What it does
render_image	Generates an image, waits for it to finish (up to about 4 minutes), and returns an inline preview plus the final image URL and a link to the file in your workspace. Also the edit tool: pass a previous render as a reference to iterate on it.
upload_image	Uploads a local photo or screenshot (up to 8 MB) to your workspace and returns a public URL to use as a render_image reference. Costs nothing.
render_video	Generates a video clip, optionally animating reference images. Returns the clip, or a generationId to collect it with get_render when it runs long.
render_3d_model	Turns one to four images of one object into a textured 3D model (GLB) on the Pro and Max plans. Returns the GLB, a lighter copy for big files and a preview, or a generationId to collect with get_render.
get_render	Fetches a render by its generationId - how you collect a video or a 3D model that was still processing. Costs nothing.
list_models	Lists the available image, video and 3D models with their categories, costs, and supported parameters.
get_credits	Reports the remaining balance and what the plan can render - image and video resolution ceilings, pro renders left this month, whether 3D models are included, and whether output is watermarked.

You rarely call these by name. Ask for what you want ("render a 16:9 image of ...") and the assistant picks the right tool and parameters.

render_image parameters

Parameter	Default	Notes
prompt	required	Plain-language description of the image.
modelId	best base model	An id from list_models. Only set it when you want a specific model.
aspectRatio	1:1	For example 16:9, 4:3, 9:16. Must be supported by the model.
resolution	1024	Long-edge scale, for example 1024 or 2048.
ingredientUrls	none	Up to 4 reference image URLs for image-to-image: a previous render to iterate on, or a URL from upload_image. URLs are validated before anything is charged.
background	none	transparent returns a PNG with a real see-through background, for cutouts, logos, icons and product assets. Works with quality pro (the default) and ultra; with fast the call is refused before anything is charged. Your assistant sets it when you ask for a transparent background. See Transparent backgrounds.
projectName	Claude MCP renders	The Visualizee project the render is saved into, found or created by name. One project per workstream (a client, a campaign, "Blog posts"), shared by all the jobs inside it.
chatName	none	The chat inside that project which groups this job's renders (a blog post title, "Kitchen", "Wheels"), found or created by title. Renders filed into a chat show up there as a conversation you can continue with Vizzy.
folderName	workspace root	An optional folder path for the project, "/"-separated for nesting (for example "Clients/Acme"), found or created on the fly.

Editing and iterating

There is no separate edit tool - editing is a render with the previous image as a reference. To change something about a result, ask for another render with the previous render's URL in ingredientUrls and an instruction that states what to preserve and what to change:

"Take the last render and change only the wheels to black 19-inch rims - keep the car, camera angle, and lighting exactly the same."

Two things make this reliable:

• State a preserve/change contract. A reference image alone drifts; naming what must stay identical ("keep the composition and materials, change only the sky to dusk") is what keeps a series consistent.
• Chain result URLs. Each result reports its URL; feed it into the next call to build a series of variations from the same base.

Your own photos work the same way. If the image lives on your machine ("here's a photo of my kitchen"), the assistant calls upload_image first, then uses the returned URL as an ingredient. Ingredient URLs are checked for reachability before a render starts, so a dead link fails fast and never costs anything.

Rendering video

render_video generates a clip, optionally animating reference images you pass in ingredientUrls. It defaults to Seedance 2.0 at 1080p (or the best resolution your plan allows), 16:9, the model's shortest duration, and audio on.

Video parameters

Parameter	Default	Notes
prompt	required	A director's brief plus a beat timeline. See below.
duration	model minimum	Seconds, from the model's supported list (Seedance 2.0: 4-15, Seedance 2.5: 4-30, Kling v4: 3-30, MiniMax H3: 5-15). Cost scales linearly.
resolution	1080p	720p, 1080p, or 4K. Capped by your plan - see below.
aspectRatio	16:9	For example 9:16 for vertical, 21:9 for cinematic.
audio	true	Generates a soundtrack. Costs nothing extra on Seedance 2.0, and is ignored by Kling v4 and MiniMax H3, which always generate audio.
ingredientUrls	none	Reference images to animate (up to 9 on Seedance 2.0 and MiniMax H3, 10 on Kling v4, up to 50 on Seedance 2.5).

Video resolution is capped per plan: Hobby renders at the lowest rung (720p, or 768p on MiniMax H3), Pro up to 1080p (2K on MiniMax H3), and Max up to 4K (1080p on Kling v4 and Seedance 2.5, which stop there). Asking for more than your plan allows names the resolution you can use rather than silently downgrading the clip - the What your plan can render table under Cost and limits has the full picture.

Collecting a finished clip

Video takes noticeably longer than an image - commonly two to eight minutes. render_video waits about two and a half minutes; if the clip is not ready by then it returns a generationId and status pending. Your assistant then calls get_render with that id to collect the finished video. The clip is charged once, and the render completes in your workspace either way.

Prompting for video

Video models respond to plain-language directing, not parameter lists. A good prompt has two parts:

1. The brief - what the subject and scene do (the reference images already show what things look like), one primary camera move, and the atmosphere.
2. A beat timeline covering the whole clip, one line per beat: [00:00-00:02] slow dolly forward begins, wind stirs the grass.

Every image you attach goes into one reference list tagged @Image1, @Image2, and so on. There is no first-frame/last-frame slot - several images are read as references to the same subject unless you explicitly ask for a transition between them.

When audio is on, describe the sound in the prompt: ambient tone, the diegetic effects the scene would make, and optionally music by character rather than by artist name. Ask for dialogue only if you want it, keep lines short, and say who is speaking. When audio is off, the prompt should not mention sound at all.

Making 3D models

render_3d_model turns one to four images of one object - a game prop, a character, a product, a piece of furniture - into a textured 3D model you can drop into Unity, Godot, Unreal, Blender or a web page. It is on the Pro and Max plans.

Parameter	Notes
images	1 to 4 views, each { url, role? }. The URL is a render URL or one from upload_image. role is front, left, back or right: give it on every image or on none, in which case the order is front, left, back, right. One view must be the front.
quality	pro (the default, Hunyuan 3D, the most detail, 900 credits) or fast for Standard (Tripo, 300 credits, smaller files). ultra renders on Pro.
modelId	A 3D model id from list_models, only when you want a specific model.
summary	What the model is of, shown in the project chat. Optional.
projectName, chatName, folderName	Where it is filed, as for render_image.

There is no text-to-3D: every model starts from an image. Ask for a 3D model of something that has no image yet and your assistant renders a concept first with render_image - one object, whole, on a plain background - then builds the model from it. A whole room or scene is not one object, so the assistant renders the piece you want on its own first.

One view works, but the model has to guess every side it cannot see. A front, a side and a back view of the same object, at the same size and in the same light, give a model whose back matches what you drew. Your assistant can make those views from an accepted render with render_image before calling render_3d_model.

A model takes 2 to 4 minutes. The tool waits about 4 minutes; if the model is not ready by then it returns a generationId, and get_render collects it. The result has the full GLB (textures up to 4K on Pro), a lighter GLB with 2K textures when the full file is over 8 MB, and an inline preview. The model is also filed in your workspace, where you can turn it and check its shape in Clay.

Where renders end up

Every MCP render is saved into your Visualizee workspace, not just returned as a URL. The workspace nests folder > project > chat > renders, and three parameters control where a render lands:

• projectName picks the project: one shared canvas per workstream, such as a client, a campaign, a car build, or "Blog posts". A project holds many unrelated chats, so your assistant should reuse one project for a whole workstream rather than creating a new one per job.
• chatName picks the chat inside that project: one per job, such as a blog post title, "Kitchen", or "Wheels". Every render and iteration for the same job goes into the same chat, and the project's chat list ends up reading like a table of contents. Each chat opens in the app as a conversation you can carry on with Vizzy.
• folderName is an optional folder path for the project, nested with "/" (for example "Clients/Acme"). Folders only group projects, so most workstreams need none.

With no names given, renders collect in a project called Claude MCP renders at the workspace root. Names match case-insensitively and are created on the fly, so repeated renders for the same job always land in the same place, even when several run in parallel. The server tells your assistant this model up front, so it picks sensible names on its own when you do not spell them out.

Each result includes a direct link to the project, which opens on the chat the render was filed into, so you can jump from your assistant into the workspace and keep iterating there.

Cost and limits

• Renders count against your plan the same way as in the app; the cost and remaining balance are reported with every result. See plans and renders for what your tier includes.
• If a render fails or cannot start, its cost is refunded automatically.
• Pro-tier models respect your plan's monthly pro-render limit. Seedance 2.0, the default video model, is a pro-tier model.
• Video is billed per second and per resolution, so clips cost hundreds to thousands of credits where an image costs 80 to 400. Read the Rendering video section above before turning an assistant loose on it.
• 3D models cost a flat 900 credits on Pro and 300 on Standard, the same for one view or four, on the Pro and Max plans only.
• For scale: a 1K image costs 80 credits on Base, 150 on Pro and 300 on Ultra. Base costs 107 at 2K, Pro 300 at 4K, and Ultra 400 at 2K and 500 at 4K. Free starts with 400 credits; Hobby includes 7,500 a month, Pro 19,250 and Max 48,000.

What your plan can render

Every request is checked against your plan before anything is charged, and a request that exceeds it comes back with the nearest thing you can run rather than a bare refusal:

Limit	Free	Hobby	Pro	Max
Image resolution	1024	2048	uncapped	uncapped
Video resolution	720p (768p on MiniMax H3)	720p (768p on MiniMax H3)	1080p (2K on MiniMax H3)	4K
3D models	no	no	yes	yes
Pro renders / month	1	15	no separate cap	no separate cap

So asking for a 2048px image on Free answers "the Free plan renders images up to 1024px - re-run with resolution 1024"; asking for 4K video on Pro points at 1080p; and exhausting the pro-render quota names a base model that does the same job without touching it. When a clip costs more than your balance covers, the reply names the largest clip your balance does cover - for example "the largest clip your balance covers is 6s 720p for 1,824 credits".

Call get_credits to read all of this up front - your balance, both resolution ceilings, pro renders left, whether 3D models are included, and whether output is watermarked - so an assistant can choose workable settings before spending anything.

Example prompts

Things you can ask once the server is connected:

• "Render a 16:9 hero image of a Scandinavian living room at golden hour and save it in project 'Blog posts', chat 'Autumn palettes'."
• "List the Visualizee models and pick the cheapest one that supports 21:9."
• "Use this product photo as a reference and render it in a minimalist concrete showroom, 4:3."
• "Take the render you just made and change only the wall color to sage green - keep everything else identical."
• "Animate that render into a 5-second clip with a slow push-in and soft room tone - tell me what it will cost first."
• "Make a 3D model of that chest for Unity - render a side and a back view first so the back is right."
• "What can my plan still render this month?"

Next