Intangible MCP · Beta
Intangible speaks MCP, so any agent that does can drive it - Claude Code, Claude Desktop, ChatGPT and the rest. Describe 3D scenes, shots and visuals in your client app. Watch it work side-by-side in a browser with Intangible. Take the wheel to fine tune results.
The agent places 3D assets in Intangible’s Build Mode, to create a virtual set based on your brief. Reasoning in 3D is (still) slow for AI so grab a coffee while it works.
Ask for specific beauty shots (or a set). The agent sets cameras and takes shots that show up in Intangible’s Compose Mode as a named list.
Ask to turn the framed shots into rendered images with art direction details. The agent hands off to Visualize Mode and generates them. Video generation coming soon.
You brief the agent in your LLM client; the 3D set takes shape in a browser window with Intangible. Keep both in view and you can watch the scene build.
3D placement is hard for an LLM. Expect objects to land roughly right so you can nudge them into place on screen (faster and fewer tokens than describing it to a prompt box).
MCP Beta Setup
One remote server, and any agent that speaks MCP can use it. Steps for the three most common clients are below, copy-paste ready.
This Beta release launches with Build-Compose-Visualize modes available to your agents. We’ll open up more functionality in Intangible to agents in the coming weeks.
https://mcp.intangible.ai/mcp
Before you start
In Claude Desktop: Settings › Connectors, then Add custom connector at the foot of the section.
On Team and Enterprise plans this is an Owner-only step, done once for the whole organization at Settings › Organization settings › Connectors. Everyone else picks up at step 3.
Name: Intangible. Remote MCP server URL:
https://mcp.intangible.ai/mcpClick Add.
Click Connect on the Intangible connector. Your browser opens, sign in to Intangible and approve. Back in Desktop, the connector shows Connected.
In a new chat, open the tools menu (the sliders icon under the message box) and make sure Intangible is toggled on.
Before you start
An owner or admin enables it for the workspace first: Workspace Settings › Permissions & Roles › Connected Data, then the developer mode toggle. Each user then switches it on for themselves at Settings › Apps & Connectors › Advanced settings.
If you cannot see the toggle at all, the workspace has not allowed it yet.
Go to Settings › Apps & Connectors and click Create. Name: Intangible. Server URL:
https://mcp.intangible.ai/mcpTransport is Streamable HTTP. Authentication is OAuth. ChatGPT will warn you that custom connectors are not verified by OpenAI, which is expected here.
Your browser opens. Sign in to Intangible and approve. The connector comes back showing its tool list.
An owner publishes from Workspace Settings › Apps › Drafts. Review the write actions on the way through. Once published it appears in every member’s Apps list, labelled custom, and each person authenticates for themselves.
On Business, a published app cannot be edited. Changing the tool list means recreating and republishing it.
Start a new chat, open the apps menu under the message box, and switch Intangible on. Confirm the tool list loaded. If it is empty, open the connector in settings and choose Refresh.
ChatGPT asks you to approve each write action before it runs, so expect a confirmation prompt the first time the agent places anything.
Before you start
Any directory works. --scope user makes it available in every project, not just one folder.
claude mcp add --transport http --scope user intangible https://mcp.intangible.ai/mcpclaude mcp listYou should see:
intangible: https://mcp.intangible.ai/mcp (HTTP) - ! Needs authentication“Needs authentication” is expected at this point.
Fully quit any running session and start it again with claude. New servers only appear after a restart.
Inside Claude Code, run the command below, select intangible, then sign in and approve in the browser window that opens.
/mcpStatus flips to Connected and the tools load. Tokens are handled for you, there are no API keys.
The wrong URL is the single most common cause of failure, so copy it from the field above rather than retyping it. Claude Code, Claude Desktop and ChatGPT each keep their own MCP config, so installing in one does not carry over to the others.
Then say hello
Send the agent a first message. hello works. So does a full brief: build me a medieval street with 10 buildings. Before it places a single vertex, the agent tells you how the collaboration works: the three Studio modes it works through, roughly how long a build takes, what it handles, and what stays in your hands.
The shape of a session. You brief the agent in your client; the scene is built and finished in Studio. The agent takes on the slow parts, standing up a large 3D scene, composing shots against it, setting camera moves. A large scene takes minutes to build and minutes more to shoot. The viewport stays empty for the first couple of minutes while the agent searches for assets, which is normal.
What you finish yourself. The agent gets a scene about 80% of the way there and hands back a named list of placements to nudge. Objects land close to right rather than exactly right, and moving one on screen yourself is faster and more accurate than describing it to a prompt box. The split is deliberate.
One tip from watching people use this: once the project is open, put the Studio window and your client side by side. The build lands in Studio while the conversation runs in the client.
Claude Code: restart it fully. Servers added while a session is running don’t appear until the next launch. Then check that claude mcp list shows the entry.
Claude Desktop: quit and reopen the app, then check Settings › Connectors. If the connector isn’t there, re-add it from step 1.
Finish the sign-in in the browser tab that opened, then return to your client. If the tab never opened, pop-ups may be blocked for the auth window.
Still stuck: remove the server and start over. A stale token from a previous attempt is the usual cause.
The agent asks you to open a project when it needs one, so follow the prompt it gives you. If you have already opened one at studio.intangible.ai, bring that tab into focus and ask the agent to try again by giving it the project URL - if it only just loaded, give it a few seconds to register.
This is a plan limit, not a setup mistake. Nearly every Intangible tool writes to your scene, and OpenAI is rolling out write actions for custom connectors on Business, Enterprise and Edu only. On personal plans the connector attaches and lists the tools, then declines to write.
If you are on a personal plan, use Claude Code or Claude Desktop instead.
Claude Code: claude mcp remove intangible
Claude Desktop: Settings › Connectors › Intangible › Remove. Then repeat the steps for your app from the top.
Two minutes of setup, then a brief. The agent stands up the scene, you finish the shot.
← Back to home