AZTechnology
한국어
BlenderBlender and AI

Using Blender MCP with ChatGPT: custom MCP servers in ChatGPT on the web, secure tunnels, and how Codex differs

What to know when you want to use Blender MCP with GPT: why the official Blender MCP cannot plug straight into ChatGPT on the web, the two routes OpenAI's docs describe (a public HTTPS address or Secure MCP Tunnel), and how they compare with Codex, the easiest way to drive Blender with a GPT model.

AZTechnology11 min read

한국어English

Based on: OpenAI's official docs (Add custom MCP server, Secure MCP Tunnel, Connect and test your plugin), Blender Lab MCP server 1.0.3, Blender 5.2 LTS. Checked in November 2026. The figures are OpenAI's official doc pages and a real PowerShell screen. The ChatGPT steps follow the official docs only.

People often ask how to use Blender MCP "with GPT". Here is the short answer. ChatGPT on the web cannot start the Blender MCP server installed on your PC. ChatGPT runs on OpenAI's computers, and the Blender MCP server only answers inside your own PC. Joining the two needs a public address or a secure tunnel, and that is a developer setup.

The easiest way to drive Blender with a GPT model is Codex. Codex is OpenAI's coding agent. It runs on your PC, so like Claude Code it can start the Blender MCP server directly.

This guide explains why ChatGPT on the web does not connect, the two routes OpenAI's docs describe, and how to choose between them and Codex. We did not try the ChatGPT side ourselves. All ChatGPT steps here are taken from OpenAI's official docs.

1. The reason it does not plug in: where the server runs

There are two ways to run the Blender MCP server.

ModeWho starts the serverWho can reach itApps that use it
Standard input/output (stdio)The AI app starts it on your PCThat one AI appClaude Code, Codex, Kimi Code CLI, Antigravity CLI, Claude desktop app
Web (HTTP)You start it yourself and leave it runningApps that come to its addressApps that run in a browser (for example the llama.cpp web UI)

ChatGPT on the web cannot use the first mode. ChatGPT runs on OpenAI's computers, not yours, so it cannot start a program on your PC.

The second mode does not work as is either. By default, the official server's web mode only answers on 127.0.0.1, which means "this PC itself". It is like a home intercom. It works inside the house, but nobody outside can call the intercom number. OpenAI's computers are outside, so they cannot reach it.

2. Starting the server in web mode on your PC

Starting point: a PC where you have downloaded the Blender MCP server. How to download it is in Section 5 of Installing the official Blender MCP.

You can see for yourself that the web-mode server only answers inside this PC.

Steps:

  1. Open PowerShell.
  2. Type the command below. Change the folder path to where you downloaded the server.
uv --directory C:\blender_mcp\mcp run blender-mcp --transport http --port 9191
  1. Check that the last line shows Uvicorn running on http://127.0.0.1:9191.
  2. When you are done, stop the server with Ctrl+C.

A PowerShell window. After the command uv --directory (server path under the user folder) run blender-mcp --transport http --port 9191, INFO lines read Started server process, Waiting for application startup, StreamableHTTP session manager started, Application startup complete, Uvicorn running on http://127.0.0.1:9191 (Press CTRL+C to quit) Figure 1. The Blender MCP server started in web mode. The address on the last line, http://127.0.0.1:9191, only answers inside this PC. The figure shows a server downloaded under the user folder (terminal screenshot)

StreamableHTTP refers to "streaming HTTP", one of MCP's web modes. It has the same name as one of the modes ChatGPT supports (Section 3). The mode matches; the problem is that the address is a home intercom that ChatGPT cannot reach.

This web mode is for browser apps on the same PC. A working example is in Driving Blender with AI on your own computer.

3. OpenAI's docs: adding a custom MCP server to ChatGPT

The "Add custom MCP server" page in OpenAI's developer docs describes how to add your own MCP server to ChatGPT.

The Add custom MCP server page on OpenAI Developers. An ELEVATED RISK label under the title, the Connect a custom MCP server paragraph, six steps under How to use, and at the bottom Supported MCP protocols: SSE and streaming HTTP Figure 2. OpenAI's guide to custom MCP servers. The ELEVATED RISK label at the top, six steps in the middle, supported modes SSE and streaming HTTP at the bottom (OpenAI official docs screenshot)

What the page says:

ItemFrom the docs
WhereChatGPT on the web. Workspace permissions and security restrictions (including Lockdown) apply
StepsChatGPT Plugins → plus button → Add custom MCP server → enter a name → under Connection, enter a Server URL or choose Tunnel → set up authentication → review the risk warning (I understand and want to continue) → Create as a plugin → install it from your plugin list, then use it in a conversation
Supported modesSSE, streaming HTTP
AuthenticationOAuth, no authentication, or a mix of both
How to useType @ in the prompt box and pick the plugin
Tools that change thingsTools without a read-only hint (readOnlyHint) count as write actions and ask for confirmation by default. Within one conversation you can have it remember approve or deny per tool; a new conversation asks again

That last row matters for Blender. In the official Blender server, reading tools such as the scene summary carry the read-only hint, while the code execution tool (execute_blender_code) is marked as a tool that can change the file. So in ChatGPT, a confirmation prompt appears before code runs. If you have it remember your approval in a conversation, it does not ask again there; a new conversation asks again.

The risk is real too. Connect only servers you trust. An untrusted server could take information or trick ChatGPT into using tools in unintended ways.

4. Two routes for ChatGPT to reach a server on your PC

The Connection field takes one of two choices.

RouteHowFor Blender MCP
A. Public HTTPS addressPut the server on an https:// address reachable from the internetNot recommended. You would be opening a server with a code execution tool to the internet
B. Secure MCP Tunneltunnel-client on your PC opens an outbound connection to OpenAI, takes requests, and passes them to your serverThe official way to connect without opening the server to the internet. A developer setup

Route B is like this: instead of publishing your home intercom, someone inside calls OpenAI first and picks up the requests. No door into the house is opened from outside.

The Secure MCP Tunnel page on OpenAI Developers. A description under the title, a paragraph saying tunnel-client opens an outbound HTTPS path, a yellow box saying public plugin submission is not supported, the What is an MCP tunnel section and the Use Secure MCP Tunnel when list Figure 3. OpenAI's Secure MCP Tunnel guide. tunnel-client in the top paragraph, the yellow box excluding public distribution, and the What is an MCP tunnel section below (OpenAI official docs screenshot)

What the docs list as requirements:

RequirementFrom the docs
Tunnel IDA tunnel_id from tunnel settings on the OpenAI Platform
Runtime API keyA key that tunnel-client uses
Your serverAn MCP server that tunnel-client can reach over stdio or HTTP
PermissionsTunnels Read + Manage to create a tunnel, Tunnels Read + Use to use one
AssociationThe tunnel must be associated with the ChatGPT workspace and Platform organization that will use it, or it will not show in ChatGPT
NetworkOnly outbound HTTPS from your PC to api.openai.com:443. No inbound ports

The docs also show a tunnel-client setting (--mcp-command) that starts a stdio server directly. In theory the Blender MCP server could go there. We did not test this route. It needs an API key and permissions, and in the end it connects a door that runs AI-written code in your Blender to an outside service, so we do not recommend it for first-time users.

5. The easy way to drive Blender with GPT: Codex

If you want a GPT model with your ChatGPT account and you want to work in Blender, Codex is much simpler than ChatGPT on the web.

PointChatGPT on the web + custom MCP serverCodex (CLI)
Where the server runsNeeds an address reachable from outside, or a tunnelYour PC. Codex starts the server itself
Registering Blender MCPPublic address or tunnel + ChatGPT Plugins setupOne command (codex mcp add) or one block in a settings file
What you needWorkspace permissions; for a tunnel, Platform permissions and an API keySign in with a ChatGPT account
Confirming code runsA confirmation for every write actionFollows Codex's approval settings
For a first-time userHardEasy to follow

How to register the official Blender MCP in Codex is in Working in Blender with Codex and GPT-6 Astra. Whether you use Codex or Claude Code, the Blender side (add-on and server) is the same, so once it is set up you only swap apps. App-by-app differences are in the comparison table of Choosing an AI tool for Blender MCP.

6. Picking a route

What you wantRoute
Build and edit Blender scenes with a GPT modelCodex (Section 5)
Drive your Blender from ChatGPT on a phone or another PCSecure MCP Tunnel (route B in Section 4). A developer setup; read every code confirmation carefully
Share one server with several people in a company workspaceDecide the tunnel and permissions with your workspace admin
Try it without a subscriptionDriving Blender with AI on your own computer

Practice

  1. Start the web-mode server with the command in Section 2 and check that the address on the last line starts with 127.0.0.1. Then stop it with Ctrl+C.
  2. From the table in Section 3, write one sentence about which tools ChatGPT asks to confirm before running.
  3. Write down why you want to use GPT with Blender, and pick the matching route from the table in Section 6.

Common mistakes and fixes

SymptomCauseFix
There is no Blender connector in ChatGPT on the webChatGPT cannot start the Blender MCP on your PCUse Codex, or set up a tunnel
I entered http://127.0.0.1:9191 as the Server URL and it does not connectThat address only answers inside this PC, so OpenAI cannot reach itYou need a public HTTPS address or Secure MCP Tunnel
I made a tunnel but it does not appear in ChatGPTThe tunnel is not associated with the ChatGPT workspaceAssociate the workspace and Platform organization as the docs say, and check Tunnels Read + Use
A confirmation pops up for every code runThe code execution tool is marked as a write actionExpected. Read the code and approve. Within one conversation you can have it remember the approval
The connection dropped when I closed the web-mode server windowWeb mode only answers while the window is openKeep the window open while you work

In one line

ChatGPT on the web cannot start the Blender MCP server on your PC → connecting it needs a public address or Secure MCP Tunnel (a developer setup) → for an easy way to drive Blender with GPT, use Codex. If the connection fails, go through When Blender MCP won't connect in order.

References

Back to the Blender guides

KEEP EXPLORING

Read next

EEVEE render of a grey fighter-jet model sitting at an angle on a beige floor, with a shadow to the lower rightBlender
Blender and AI11 min read

Cleaning up an AI-generated 3D model in Blender: merge split points, set the size, reduce faces

Import a fighter-jet model GLB made with Meshy into Blender 5.2, read its state from the statistics, merge split points with Merge by Distance, tidy normals and smooth shading, set the size, and cut the face count to one tenth with Decimate.

A 960x540 EEVEE render. A red cube on a gray floor with a smooth gray monkey head resting on it by the tip of its chin, and a shadow falling back to the left (Blender render)Blender
Blender and AI15 min read

Ten first prompts for Blender MCP: from reading the scene to rendering, seen through the real answers

After connecting the official Blender MCP to an AI app, we sent ten first prompts in order: read, create, edit, render, and review. This guide shows what the AI did for each one and what a person still needs to check, with screenshots.

EEVEE render of a bright gold metal monkey head on a grey floor, with sharp reflections on the brows, eyeballs, and ears and an interior photo background with beige wallpaper and a dark window (Blender render)Blender
Blender and AI14 min read

Letting AI set up lights and materials: three-point lighting and PBR over MCP, compared in renders

Through the official Blender MCP, hand the three-point lighting and the gold metal and matte floor materials of a monkey head scene to AI, read the real result in the panels, and compare three EEVEE renders (default light, the AI's result, and a hand-tuned version) to learn why metal looks dull and which values are worth changing by hand.

EEVEE render of a low-poly room with two walls and a wooden floor. A bookshelf with books on its shelves at front left, a bed with a blue blanket, a pinkish rug, a desk, chair, and desk lamp under the window, and a potted plant in the right corner (Blender render)Blender
Blender and AI13 min read

Building a low-poly room with Blender MCP: the order of questions and what to fix by hand

Send five questions in one conversation through the official Blender MCP to build a low-poly room with walls, a window, furniture, props, and lights, then use the AI's notes, a check that does not change the scene, and the first render to fix the bookshelf direction, the default light left on, and a dim desk lamp by hand.

Front view of twelve fence pickets selected with an orange outline, standing at even spacing on the red X axis, with Count 6 in the Array panel of the Modifiers tab on the right (Blender screenshot)Blender
Blender and AI14 min read

Letting AI build a modifier stack: Array, Mirror, and Bevel over MCP, then fixing it by hand

Through the official Blender MCP, ask AI to stack Array, Mirror, and Bevel modifiers on one fence picket, read the real result in the Modifiers tab, and fix the picket count the AI read differently with one number. Two common mistakes, unapplied scale and the Mirror origin, are covered too.

Two bar charts. Left: tokens for nine runs (A1 to C3), mostly input read from cache. Right: estimated cost, highest for the first run of each group and lowest for C2 and C3 at 0.065 dollars (chart)Blender
Blender and AI12 min read

Measuring Blender MCP token use: the same job three times each, and how to save

With Claude Code and the official Blender MCP, we ran 'place five cylinders' in three question styles, three times each, nine runs in all. We measured input and output tokens, time, and estimated cost, and compared with real numbers how much it helps to skip screenshot checks and to give exact values.