Connecting through MCP: Claude Code, Codex, Cursor

For AI agents and code editors. The agent runs the code, looks at the result and keeps going.

What MCP is and why it is here

MCP, the Model Context Protocol, is the common way AI agents attach external tools. Blender Bridge ships such a server inside the application, so there is nothing separate to download or build.

The difference from the extension is who presses the button. With the extension you do; with MCP the agent does. It can read the state of the scene, run code, look at the result in a screenshot and continue on its own until it matches what you asked for.

Setup

Open the Blender Bridge application and sign in.
In settings choose AI connections (MCP). It contains a ready registration with the correct paths.
Copy it into your agent — Claude Code, Codex, Cursor, VS Code or another environment that speaks MCP.
Restart the environment so that it picks up the new server.
Ask the agent to check the connection: it should see Blender and report that it is ready.

What the agent can do

The server gives the agent a set of tools: check the connection, read the state of the scene, run code in Blender, capture the view, undo the last action, cancel an operation that is taking too long. That is enough for the agent to work on its own and correct itself from what it sees.

Undo works here too: if the agent did the wrong thing, ask it to put the scene back.

If the agent cannot see Blender

Check that the Blender Bridge application is open and you are signed in.
Check that Blender is open and the add-on is enabled.
Restart the agent environment: the tool list is read at startup.
If the error mentions the pairing key, open Pairing key in the application's settings and set it again.

Frequently asked questions

How is this different from the extension, apart from setup?

In who presses the button. With the extension you run the code after seeing it. Over MCP the agent runs it itself, looks at the result in a screenshot and continues until it matches. The first gives control, the second gives speed.

What exactly can the agent do in my Blender?

Exactly what the server's tools allow: read the state of the scene, run code, capture the view, undo the last action, cancel an operation that is dragging on, check the connection. The server provides nothing beyond that.

Can I limit the agent so it does not touch anything else?

The most reliable way is to work in a separate file and save before a large change. Undo of the last action is always available, but it is built for one operation, not for ten steps back.

The agent cannot see Blender.

An environment reads its tool list at startup: if you registered the server and did not restart, it does not know about it. Then, in order: the application is open, you are signed in, Blender is open, the add-on is enabled.

Next: Which AI to use with Blender