Claude can’t draw a picture.
Ask it for an illustration and you get SVG code. Anthropic’s help center says it plainly: Claude “doesn’t generate photos or illustrations the way image-generation tools do.”
At first glance this is a real hole. ChatGPT and Gemini make images in the same chat where you write. Ethan Mollick said it this week: “Claude not having an image generator is a gap in agentic projects that involve knowledge work.” Do you have to leave Claude every time you need a slide, a mockup or an infographic?
But think about it for a minute, and the hole is smaller than it looks.
Claude Code uses tools, and a tool can be an image model. A good open one came out on September 20, 2026: Qwen-Image-2.1, from Alibaba’s Qwen team.
To connect the two, add the Hugging Face MCP server to Claude Code, then add a Space that runs Qwen-Image-2.1 as a tool. You need a free Hugging Face account. You don’t need a GPU.
Why Qwen-Image-2.1
It suits documents. A transparent PNG comes straight out of the model, with no background to remove afterwards. The text inside its pictures stays readable, which matters for diagrams and infographics. And it edits as well as generates, so you can ask for a fix to the picture you just got.
There are two ways to run it. This guide follows the first.
A Hugging Face Space
A free account and one command. The model runs on Hugging Face's GPUs, and Claude Code calls it through MCP.
Your own computer
Diffusers and ComfyUI both support the model, and Claude Code can write and run the script. One test on a Mac peaked at 31 GB of memory. See how to run it locally.
Connect Claude Code to Hugging Face
Create a free Hugging Face account, then run:
claude mcp add --transport http --scope user \
hf-mcp-server "https://huggingface.co/mcp?login"
Keep the quotes, or zsh will trip over the ? in the URL. --scope user makes the server available in every folder, not only the one you ran the command in.
Start Claude Code, type /mcp, and sign in to Hugging Face.
Add a Space that runs the model
A Space works as a tool only if it shows a grey MCP badge. The official Qwen demo doesn’t have one, so use a community Space.
Open the list of Spaces with MCP support and search for Qwen-Image-2.1. On September 21 there were two: kingboy123/Qwen-Image-2.1-Demo and aldistefanus/qwen-image-2.1.
Click the MCP badge on the Space, choose Add to MCP tools and confirm. Restart Claude Code, and the Space is a tool.
These Spaces belong to individuals, and your prompt passes through them. Keep private material out of the prompt.
Ask for a picture
An image model will draw anything you describe. The hard part is knowing which picture the page needs. Claude Code can read your draft, so start there:
Read draft.md. Suggest two places where a picture would help,
and say what each one should show.
Then ask for one:
Generate the first one with Qwen-Image-2.1: transparent background, 16:9.
Save it as images/pipeline.png and add it to draft.md where we agreed.
Then let it check its own work. Claude Code can’t draw, but it can read image files. Ask it to open the PNG and compare it with what you asked for.
Read the text in the picture yourself. One wrong letter in a diagram is the mistake a reader notices first.
How many pictures are free
Hugging Face counts GPU time per day. The first Space above requests about two minutes for a 2048×2048 picture.
| Account | GPU time per day | Full-size pictures |
|---|---|---|
| Free | 5 minutes | 2 |
| PRO | 40 minutes, then $1 per 10 minutes | About 20 |
Work at a smaller size with fewer steps, and save full size for the version you keep.
Writing a document with pictures
In a terminal, Claude Code hands you a file path. To see the picture you switch to Finder. To see where it landed in the draft you open a third app. For one icon that’s fine. For a document with ten pictures, you spend the afternoon switching windows.
We make StashBase, a writing workspace that runs Claude Code or Codex beside your documents, on the subscription you already have.
The draft, the chat and the images/ folder are in one window. Open a picture beside the chat, say what’s wrong with it, and ask for another version. The draft and the pictures stay ordinary files in your folder.
StashBase doesn’t generate the pictures. That is still Claude Code and Qwen-Image-2.1. StashBase is where you read the draft, look at the picture and ask for the next one. The agent setup guide covers installing and signing in.