Guide ยท MCP image generation (2026)
Claude cannot draw by itself; it needs an MCP server that can. Cubric Studio is a free desktop app with one built in. Connect it in one click and your agent makes images, video and GIFs with open models on your own graphics card, looks at what it made, and keeps everything in a real project.
Short answer
For MCP image generation without building a ComfyUI workflow, connect your agent to Cubric Studio: free, open source, local, with Claude Code, Claude Desktop, Codex and Antigravity one click away in Settings > Connect an agent. If you already build ComfyUI graphs and want the agent to run those, use a ComfyUI MCP server instead. Both compared below.
MCP (Model Context Protocol) lets an AI agent call tools in another program. The agent does the talking and the planning; the program does the work. For pictures, that program has to hold the models, run them on a graphics card and hand the result back. Cubric Studio is that program: it runs on your computer, and its MCP server answers only on your computer.
Everything lands as a normal card in the app's gallery, with its prompt and settings saved, so you can open it, edit it by hand or carry on in the app.
You need Cubric Studio 2.0 or newer, installed and open. The easy way, for all four: in the app, open Settings > Connect an agent and press Connect next to yours.
/plugin marketplace add https://github.com/MadPonyInteractive/cubric-studio-agents.git, then /plugin install cubric-studio@cubric-studio.codex plugin marketplace add MadPonyInteractive/cubric-studio-agents, then codex plugin add cubric-studio@cubric-studio.Cubric Studio is also listed in the official MCP Registry as io.github.MadPonyInteractive/cubric-studio. Full steps: install guide on GitHub, Connect an agent in the docs.
With local models, yes. Claude, Codex or Antigravity plan the work; Cubric Studio runs open models such as Krea 2, FLUX.2 Klein, SDXL, Qwen Image Edit, Wan 2.2 and LTX 2.3 on your own graphics card, at no cost. If your card is small, cloud models (Seedream, Nano Banana, Seedance, Veo) are there on your own DeepInfra key, and the agent must show you the price and get your yes before each paid run. There are no Cubric credits.
One thing to know: a picture you paste into the Claude chat does not reach the app as a file. Give the agent the picture's path on your disk, or a card already in the gallery.
| Cubric Studio | A ComfyUI MCP server | |
|---|---|---|
| What the agent drives | A finished app: models, Flows, gallery | Your own ComfyUI and its graphs |
| Install ComfyUI yourself | No, the app installs and looks after its engine | Yes |
| Installing models | One tool call, after your yes | By hand, or as the server allows |
| Your own node workflows | No, a curated lineup | Yes |
| Results | Gallery cards with prompt and settings saved | Files in ComfyUI's output folder |
| A person can carry on by hand | Yes, in the same app, no nodes | In the ComfyUI graph |
Comfy's own MCP server, in beta: about 40 tools to submit and watch jobs, inspect installed models and nodes, and build workflows. It needs Python, comfy-cli and a ComfyUI you install and run. Licensed AGPL-3.0 or commercial.
A free (AGPL-3.0) native desktop app for images, video and audio on your own NVIDIA graphics card, with no Python, Docker or ComfyUI. It exposes a small MCP server (7 tools) for outside agents; there is no agent inside the app. Windows, Linux and macOS.
A free (GPL-3.0) video workstation whose MCP server, by its README, offers 100+ tools: plan shots, generate, arrange the timeline, mix audio, export. All generation needs your own running ComfyUI.
An open-source (Apache-2.0) infinite canvas with a video editor and an MCP connection (30 tools) so an agent works in the same project. It runs from source with Node and Chrome, and generation goes through the BeatAPI cloud.
A small personal MCP server (MIT) that turns image generation, editing and video into simple tool calls over your own ComfyUI. Two models, weights installed by hand, no releases yet.
Facts from each project's own GitHub page, checked 22 October 2026. Tool counts are the projects' own figures.
Hosted generation services offer MCP servers too; those generate on their own machines and bill per run.
Cubric Studio has its own agent, Cosmo, in every project, with the same tools. The AI image generator with an agent built in.
Not by itself. Claude writes text and code; it makes images and video by calling a tool over MCP. Connect Cubric Studio and Claude Code or Claude Desktop can make images, video and GIFs with the models installed on your computer, see the results and carry on.
Yes, with local models. Cubric Studio is free and open source, and its local models run on your own graphics card at no cost. Paid cloud models are optional, bill your own DeepInfra key, and the agent must show you the price and get your yes before each one.
For image and video generation on your own computer with no ComfyUI workflow to build, Cubric Studio: one click to connect, models installed for you, results land as gallery cards with their prompts saved. If you already build ComfyUI graphs and want your agent to run them, a ComfyUI MCP server is the better fit.
Install Cubric Studio 2.0 or newer, open it, then go to Settings > Connect an agent and press Connect next to Claude Code. Restart Claude Code. The plugin can also be added by hand from the MadPonyInteractive/cubric-studio-agents marketplace.
Cubric Studio's MCP server runs locally and drives a ComfyUI engine it installs and looks after itself, so your agent never has to build a workflow. Comfy's own MCP server and community servers expose a ComfyUI you run yourself, including your own graphs.
Yes. Through Cubric Studio an agent can make video from a prompt or a picture, run Flows such as Extend Video, and make and edit GIFs. A long job returns at once and the agent checks back, so the chat never freezes.
Not from the browser: the ChatGPT and Gemini web apps run in the cloud and cannot reach an app on your computer. Use Codex, which comes with a ChatGPT plan, or Google Antigravity, which works with a Google account.
No. None of Cubric Studio's MCP tools deletes a card, a project or a file. The one delete is a saved routine. The server answers only on your own computer.
Yes. Cubric Studio is free and puts no limit on what you do with what you make: it is yours. Each model comes with its own licence from the people who made it. Most allow commercial use; a few are for non-commercial use only, and some ask larger companies for a separate agreement with the model's creator. Where a model has terms you must accept, the app shows them before it downloads.
Cubric Studio is free and open source. Unzip it and run it: no installer, no account.
Windows, macOS and Linux.