Desktop 0.4.1 (preview): Download · Getting started · Desktop build and authentication

Windows x64 EXE, Apple Silicon DMG, and Linux x64 AppImage / deb. Existing web / CLI documentation follows. Live AI requests require the user's own authorized account.

Chat2Blend

Describe it. Watch Blender build it.

Turn ChatGPT into a live 3D modelling partner for Blender.

No copy/paste. No OpenAI API key. No second coding agent.

npm install -g chat2blend

Get started · How it feels · How it works · 中文

---

What does it feel like?

You type this into the Chat2Blend CLI:

c2b brain "a modern three-seat fabric sofa" --wait

ChatGPT writes the Blender Python. Chat2Blend runs each finished chunk in Blender while the rest is still being generated. You sit back and watch the sofa appear — arm by arm, cushion by cushion.

You
 │  "a modern three-seat fabric sofa"
 ▼
ChatGPT  ─────  writes Blender Python
 │
 ▼
Chat2Blend  ───  streams finished chunks
 │
 ▼
Blender  ──────  executes them live
 │
 ▼
🛋  a sofa in your scene

Below is a real, unedited screenshot — a genuine Blender GUI run, with 6 streaming chunks creating 14 objects.

_{Real output of a modern three-seat fabric sofa with rounded cushions, streamed into a visible Blender 4.2.9. The outliner on the right lists the 14 C2B_* objects Chat2Blend created — Arm_L/R, Back_1..3, Seat_1..3, Leg_1..4, Frame_Base/Back. This run used synthetic chunks (scripts/e2e-blender.mjs) to isolate the Blender execution path.}

_{For the full round trip through a real ChatGPT session, see the numbers in docs/PROJECT_STATUS.md: "a low-poly wooden side table" → 8 chunks, 38 objects, first geometry visible in 38.6s.}

---

Why?

Generating Blender Python with an LLM already works. The annoying part is everything around it.

Normally you:

1. Ask ChatGPT for bpy code
2. Wait for the entire answer
3. Copy the code
4. Open Blender
5. Paste it into the script editor
6. Run it
7. Hit an error
8. Go back to ChatGPT
9. Repeat

Chat2Blend removes the middle.

You describe the asset. ChatGPT writes the Blender code. Completed chunks run in Blender the moment they arrive, so you see progress instead of waiting for a wall of text.

• No copy/paste. The code never touches your clipboard.
• No API billing. It uses the ChatGPT desktop app you already pay for.
• No second agent. Your coding agent doesn't burn tokens rewriting bpy.
• Visible execution. Everything happens in the Blender window in front of you.
• Local-first. Loopback only. No tunnels, no cloud, no telemetry.

---

⚡ Quick start (3 minutes)

1. Install

npm install -g chat2blend

Requires Node.js ≥ 20 and Blender 4.x.

2. Run setup

c2b setup

This checks your machine and prints the exact next steps for your install — including the absolute path of the Blender add-on you need to pick.

3. Install the Blender add-on

In Blender: Edit ▸ Preferences ▸ Add-ons ▸ Install…

Choose the path printed by c2b setup (it ends in blender_addon/chat2blend/__init__.py).

Then enable Chat2Blend and click Connect in the Chat2Blend tab of Blender's side bar (N).

4. Start the bridge and attach ChatGPT

c2b start          # starts the local bridge
c2b brain-attach   # finds (or launches) the ChatGPT desktop app

The ChatGPT desktop app must be installed and logged in — c2b doctor will tell you.

5. Build something

c2b brain "a low-poly wooden side table" --wait

Watch your Blender window. That's it.

Stuck? Run c2b status at any point. It reports whether the bridge, Blender and the ChatGPT app are each ready, and tells you what to do about the one that isn't.

---

Examples

c2b brain "a low-poly wooden side table" --wait
c2b brain "a modern three-seat fabric sofa" --wait
c2b brain "a stack of three ceramic coffee mugs" --wait
c2b brain "a mid-century modern armchair" --wait

Don't want to wait for it? Drop --wait and poll instead:

c2b brain "a wooden bookshelf with 4 shelves"
c2b brain-status <jobId>

Other ways in:

c2b exec examples/cube.py     # send a Python file straight into Blender
c2b prompt "a modern sofa"    # print the modelling prompt Chat2Blend uses
c2b jobs                      # list recent jobs
c2b logs                      # tail the bridge log

---

How it works

ChatGPT Desktop App  (your login, no API key)
      │  CDP on 127.0.0.1:9333
      ▼
Local Brain  ── finds the window, fills the composer, reads the code blocks
      │  HTTP on 127.0.0.1:8787
      ▼
Bridge  ── jobs, queue, auth, logs
      │  line-delimited JSON over TCP on 127.0.0.1:8788
      ▼
Blender Add-on  ── executes on the main thread, shared namespace
      ▼
Blender Scene

The local brain is zero-dependency TypeScript that talks to the ChatGPT desktop app over the Chrome DevTools Protocol. It fills the composer, waits for the answer, and extracts Python code blocks from the rendered conversation. No browser extension, no scraped DOM, no third-party cookies.

The bridge is zero-dependency Node.js — /api/* routes plus a line-delimited JSON socket to Blender.

The Blender add-on is pure Python stdlib + bpy. Networking runs on a background thread; execution is drained on the main thread via bpy.app.timers, which is why your Blender window stays responsive and visible.

The C2B chunk protocol

To get streaming execution, the LLM wraps its output in chunk markers:

# C2B:CHUNK arm_left
...python...
# C2B:END

Rules: Blender 4.x bpy API; helpers defined before use; earlier chunks cannot depend on later chunks; every chunk is a standalone valid Python block.

If the model doesn't follow the protocol, Chat2Blend falls back to executing complete ```python fences after generation ends.

See docs/PROTOCOL.md and docs/STREAMING.md.

---

Agent integration

Coding agents should not generate large bpy scripts when Chat2Blend is available.

Use Chat2Blend to generate the Blender asset. Run `c2b status` first; if Blender is
not connected, tell the user to open Blender and enable the add-on. Then run
`c2b brain "<asset description>" --wait` and report the resulting job status.

The agent never writes or sees the generated bpy code — it only receives compact chunk status, object counts and time-to-first-frame.

See skill/SKILL.md and docs/AGENT_INTEGRATION.md.

---

Current support

Platform	Status
Windows 10/11	✅ Tested end-to-end (Blender 4.2.9 + ChatGPT desktop app)
macOS	🧪 Architecture supported, not yet verified on real hardware
Linux	🧪 Architecture supported, not yet verified on real hardware

The core loop is real and verified: natural language → ChatGPT desktop app → bridge → Blender GUI → 3D model.

---

Security

Chat2Blend executes LLM-generated Python inside Blender. That is inherently powerful, so:

• The bridge binds 127.0.0.1 only. Binding 0.0.0.0 is refused.
• The local brain only connects to the ChatGPT desktop app on this machine.
• Web origins outside http://127.0.0.1 / http://localhost are rejected.
• We never ask for ChatGPT credentials, OpenAI API keys, or browser cookies.

Read the full threat model in docs/SECURITY.md. To report a vulnerability, see SECURITY.md.

---

Development

git clone https://github.com/nanhudev/chat2blend.git
cd chat2blend
npm install
npm run build          # TypeScript + extension
npm run typecheck
npm test               # parser + bridge integration tests
npm run package:blender

End-to-end verification

A real GUI test that drives the ChatGPT desktop app:

node scripts/e2e-brain.mjs "a low-poly wooden side table"

It starts a bridge, opens the Blender GUI, connects the add-on, drives the local ChatGPT app, waits for the generated chunks and checks that the model appears in the scene. Requires a logged-in ChatGPT desktop app.

A synthetic, no-LLM variant that streams chunks from disk:

node scripts/e2e-blender.mjs examples/sofa_chunks.py 1200

There is a hard rule in this project: no fake success. See CONTRIBUTING.md.

---

Roadmap

• Multi-provider adapters (Claude, Gemini, DeepSeek, Grok, local LLMs)
• Manual edit / retry / skip of individual chunks
• Headless mode for CI
• Godot export pipeline

See docs/ROADMAP.md.

---

Documentation

Doc	What it covers
docs/ARCHITECTURE.md	Component-by-component design
docs/PROTOCOL.md	The C2B/1 chunk protocol
docs/STREAMING.md	Streaming parser and dedupe
docs/AGENT_INTEGRATION.md	Using Chat2Blend from a coding agent
docs/SECURITY.md	Threat model
docs/ROADMAP.md	What's next
docs/PROJECT_STATUS.md	Verified vs. experimental
CHANGELOG.md	Release history

---

Contributing

Bug reports, ideas and PRs are welcome — start with CONTRIBUTING.md. Please also read the Code of Conduct.

---

License

MIT © Chat2Blend contributors

Chat2Blend is an independent open-source project. It is not affiliated with, endorsed by, or supported by OpenAI or the Blender Foundation. "ChatGPT" is a trademark of OpenAI; "Blender" is a trademark of the Blender Foundation. You must comply with the terms of any chat subscription you use this with.