Marso MCP: 3D Assets and PBR for Your AI Agent (2026)

Marso MCP is the MCP server for Marso Studio. It connects the AI agent you already use, such as Claude, ChatGPT, Codex or Claude Code, to Marso Studio, and through the Marso plugins to Blender and Isaac Sim. You add one server URL, sign in with your Marso Studio account and ask for the result in a sentence. The agent runs the steps, prices the run before it spends a credit, and saves the work to a project you open on marso.app.

What does one request to Marso MCP look like?

One request to Marso MCP looks like this. It's the message from the video above, typed into ChatGPT:

Stock this garage with props my robot can practise on. Put them in Isaac Sim.

From that one message, the agent:

Generates an image of a garage in Marso Studio.
Finds the objects in it. In this run it found 42 and kept the 12 a robot can handle.
Makes a reference, a 3D model and predicted PBR materials for each of the 12, as one batch.
Puts the finished assets on an Isaac Sim stage, where a robot with a suction gripper practises on them.

Every step is a node in a Marso Studio project, so the run is still there when the chat ends. Open it on marso.app and you can see what the agent did, change a step and run it again.

Why does a scene's worth of props take so long?

A scene's worth of props takes so long because every prop is the same run of small steps. You find or make a reference image. You generate the mesh. You get its materials. You check what it will cost, and you keep track of which of the forty barrels, crates and bottles are done.

None of those steps is hard on its own. The problem is the count. With dozens of props, the work stops being craft and becomes bookkeeping, and the bookkeeping is what eats the afternoon.

More and more artists and technical teams already plan and script that kind of work in an AI agent. Until now, using Marso Studio meant leaving that conversation and doing each step by hand in the app. Marso MCP closes that gap. You imagine, your agent executes.

What is Marso MCP?

Marso MCP is a remote server that lets an AI agent work in Marso Studio on your behalf. MCP, the Model Context Protocol, is the open standard that agents such as Claude and ChatGPT use to connect to outside tools. Marso MCP speaks that standard, so your agent can drive Marso Studio the way it already drives your files or your design tools.

The names matter, so here they are plainly. Marso MCP is the connection. Marso Studio is still the app, and still the account you sign in with. Everything the agent makes lands in Marso Studio, where you can see it, change it and keep it.

There is nothing to install. The server address is:

https://mcp.api.marso.app/mcp

It has setup steps for Claude, ChatGPT, Codex and Claude Code. Any other client that connects to remote MCP servers with OAuth sign-in, such as VS Code or Gemini CLI, uses the same URL.

What can your agent do in Marso Studio?

Your agent can run six workflows in Marso Studio through Marso MCP. Each one comes with a prompt you can give your agent word for word.

Every object in a scene. Start from a photo or a generated scene. The agent finds the objects, you pick the ones you want, and each becomes a mesh with predicted PBR materials. Try: "Find the objects in this kitchen photo and make each one a 3D asset with PBR materials."
One object, start to finish. The reference, the mesh and the materials, from a description or a reference image. It also makes sets with one look, and versions such as old, wet or broken. Try: "Make a weathered oak barrel with iron bands as a 3D asset with PBR materials."
A set from one idea. No image needed. Each item in the set becomes its own mesh with predicted PBR materials. Try: "Make 24 unique potion bottles as 3D assets with PBR materials."
One change across a set. One change, such as a new material or finish, applied to a whole group of images at once. Try: "Make all of these images weathered brass."
Materials for your meshes. Predicted albedo, roughness, metallic and IOR maps, with a USDZ, for a textured GLB, USDZ or FBX. Try: "Make PBR materials for these three meshes."
Maps from a flat image. The predicted albedo, roughness, metallic and IOR maps of a texture or a swatch. Try: "Turn this fabric swatch into a PBR material."

The potion bottles are the point of the whole thing. Twenty-four bottles is twenty-four runs of the same steps, and you asked for it once.

The materials come from image-to-material (I2M), our best PBR model, which is trained on material data we measure ourselves.

What does it look like in the chat?

Results show up in the chat as cards, so you can follow the work without switching to another window.

A 3D viewer, so you can turn the asset over and look at the materials before you open anything.
The detected objects, with a box over each one and a tick next to each, so you choose what gets made before anything is built.
Live progress while a run is going.

The detection card is the part we care about most. Generating assets is the easy half. Deciding which twelve of the forty things in a scene are worth building is the half that decides whether a tool survives contact with a real project, and that decision stays with you.

How does Marso MCP keep spending in your hands?

Marso MCP keeps spending in your hands with three controls between the agent and your credit balance.

A price before anything runs. The agent checks the current price and your balance, and quotes what the whole run will cost before it spends a credit.
An allowance you set. Tell the agent how many credits a run may spend without asking. Runs under it go ahead, and runs over it stop and wait for you.
No double charges. If the agent tries to repeat a paid run while the first one is still pending, the repeat is refused, so the same work is never charged twice.

That matters more with an agent than with a person clicking buttons. An agent is quick to retry and happy to scale up a request, and a spend model that assumes a human is watching every click does not hold once nobody is.

The server itself is free to connect. Runs use the same credits as the Marso Studio web app, from the same balance. Steps that only build the graph are free, and steps that run a model cost credits.

Where does the work end up?

The work ends up in your Marso Studio account. A project run lands on a node graph, with a column for each stage and a row for each object, in a project you open on marso.app. It's the same kind of graph you would build by hand, so you can change a node, swap an input and run it again.

That is the difference between an agent that hands you a file and one that hands you a workflow. A file is the end of the conversation. A node graph is somewhere to pick up from.

Images and meshes you upload go to your Marso library. Material runs on your own files work slightly differently: the maps come back as download links in the chat, and they are not stored in a project.

The agent also can't delete your work. The server has no tool that deletes a node or a project. An agent can turn a node off, which you can undo, and anything you want gone you delete yourself on the canvas.

Does Marso MCP work with Blender and Isaac Sim?

Yes. Marso MCP drives both apps directly, through the Marso plugin for each one. Switch on Connect to Marso Agents in the plugin, and the same conversation that builds a set in Marso Studio can work in the scene you have open.

In Blender (Marso add-on 1.4.1 or later, Blender 5.2 or later), your agent can:

Check the scene with a viewport screenshot.
Put Marso assets into the scene, and add PBR materials to your own meshes.
Send the viewport, a camera render or a mesh to start a generation.
Set up a studio with backdrops, framing, lighting and HDRIs.
Make turntables and camera paths, and render a still or a movie into the chat.

In Isaac Sim (Marso Library extension 0.2.0 or later, Isaac Sim 6.0 or later), your agent can:

Put finished assets on the stage as USD, ready for physics.
Film physics tests for real-world scale, drops, friction, wind and collisions.
Film robot tests, from pick-and-place and suction grips to a humanoid and a two-robot handover.
Play each video in the chat.

Each step that spends credits is priced first, and Sim Studio videos are free to run. Each plugin connects out to Marso, so nothing listens on your machine, and the agent only sees the apps signed in to your Marso account. It can run a fixed list of Marso commands, and it can't read your files, run code in the app or change the plugin's settings.

These plugin versions arrive on the Plugins page on marso.app soon.

For robotics and simulation teams, Isaac Sim is where this matters most. A simulation wants many physically plausible objects rather than a handful of hero assets, and the garage at the top of this post is that job in one message.

How do you connect Marso MCP?

Connecting Marso MCP takes one server URL and one sign-in.

Add the server. In Claude or ChatGPT, add https://mcp.api.marso.app/mcp as a custom connector named Marso. In Codex, run codex mcp add marso --url https://mcp.api.marso.app/mcp, then codex mcp login marso. In Claude Code, run claude mcp add --transport http marso https://mcp.api.marso.app/mcp, then run /mcp and pick marso.
Sign in with your Marso Studio account. The first time, the sign-in page asks for your Marso API key and shows you how to make one. You paste it there, never in the chat.
Ask for what you want. Start with one of the prompts above and adjust from there.

Each teammate signs in as themselves, so a studio can hand Marso MCP to the whole team without sharing one login. The Marso MCP page has the steps for each client.

What is it like for a technical team?

For a technical team, Marso MCP is one remote server with a standard sign-in. Nobody installs anything to use it in Marso Studio, and in Codex or Claude Code it is added from the same terminal you already work in.

The agent chains generation, detection and PBR in a single conversation, so a request like the garage above runs end to end without you stitching steps together. Slow runs return at once with a progress card rather than leaving the conversation hanging, and if a run is interrupted and resumed, the stages that already finished are skipped.

The server runs remotely and can't read files on your machine. When you hand your agent an image or a mesh, it uploads the file for you or gives you an upload link. It takes PNG, JPG and WebP images, and GLB, USDZ and FBX meshes.

Where Marso Studio fits

Marso Studio is the node-based web app at marso.app where every project Marso MCP builds lives. It takes a textured 3D asset as GLB, USD or FBX, or starts from an image and generates the mesh, and it predicts albedo, roughness, metallic and IOR with I2M. You export PNG maps or a textured USD file.

Studio output is physically-based PBR predicted by I2M, trained on measured material data. Measured PBR comes only from Marso Measure, which captures real objects. You can read how the two relate in what measured PBR is, and about the model itself in the I2M announcement.

Marso MCP does not change what Marso Studio makes. It changes who runs the steps. You describe the result, your agent does the clicking, and you approve the price and keep the project.

FAQ

What is Marso MCP? Marso MCP is the official MCP server for Marso Studio, built and run by M-XR. It connects an AI agent such as Claude, ChatGPT, Codex or Claude Code to Marso Studio, so the agent can make 3D assets with predicted PBR materials and save them to a Marso Studio project. Through the Marso plugins it also works in Blender and Isaac Sim.

Which AI agents work with Marso MCP? Any client that connects to remote MCP servers over HTTP and supports OAuth sign-in. There are setup steps for Claude, ChatGPT, Codex and Claude Code, and other clients, such as VS Code and Gemini CLI, connect with the same URL: https://mcp.api.marso.app/mcp.

Does Marso MCP work with Blender and Isaac Sim? Yes, through the Marso plugin for each app. Blender needs the Marso add-on 1.4.1 or later and Blender 5.2 or later. Isaac Sim needs the Marso Library extension 0.2.0 or later and Isaac Sim 6.0 or later. These plugin versions arrive on the Plugins page on marso.app soon.

Will the agent spend my credits without asking? Only up to an allowance you set. The agent checks the price and your balance before it runs anything, runs over your allowance stop and ask for approval, and a repeat of a paid run is refused while the first is still pending.

Are the materials measured or predicted? Predicted. I2M, our image-to-material model, predicts every material. It learned from material data that we measure from real objects with Marso Measure. Measured describes the training data, never the output.

Connect your agent

Marso MCP works with the Marso Studio account you already have. Add the server to the agent you already use, sign in, and ask for the set you have been putting off.

Connect your agent to Marso MCP

Marso MCP: 3D Assets and PBR for Your AI Agent (2026)