You asked Claude for a product shot and got a paragraph explaining that it cannot make one. So you searched for how to generate images with Claude, found a dozen guides promising workarounds, installed something, and ended up with a connector that says “failed to connect” and no picture. That loop is the most common experience people have with this question, and almost every guide skips the part that explains why it happens.
Here is the short version. Claude has no image model attached to it, and no prompt changes that. What it does have is a set of documented routes to a picture: visuals it draws itself in code, a connector to an image model that lives somewhere else, and the plain old job of writing a better prompt for a tool that does generate images. This guide walks all six, names the OpenAI and Google models you can actually reach, and flags the two settings that quietly break most setups.
Claude has no image model of its own. Anthropic’s help center states Claude does not generate photos or illustrations the way image tools do. You can still get images three ways: ask Claude for a diagram or chart it draws in HTML and SVG, connect an image generator through a connector, or have Claude write the prompt.
Can Claude generate images? What Anthropic actually says
No. Claude cannot generate a photo or an illustration from a text prompt on any plan. Anthropic’s help center says Claude does not produce images the way image-generation tools do. It can build diagrams, charts and interactive visuals in your conversation, written in HTML and SVG, and it can view and analyze images you upload.
Go to the primary source, not the blog posts
This claim is worth checking yourself, because the pages that rank for it disagree about the details. The support article “Can Claude produce images?” on Anthropic’s help center is the document that settles it. Read it before you trust a workaround that a vendor blog is selling you.
That makes this a confirmed limitation, not an expert opinion or a rumor. It is stated in the maker’s own documentation. Plenty of competing guides describe it as a temporary gap that a clever prompt gets around. There is no evidence for that, and Anthropic has published none.
What the no rules out, and what it does not
The no is narrow. It covers one thing: Claude turning text into pixels by itself. It does not cover Claude drawing with code, which it does well, and it does not cover Claude asking another program for a picture on your behalf, which is what every working method below depends on.
The limit is also not a plan problem. Upgrading from Free to Pro, Max or Enterprise adds features, and none of them is an image model. Anyone selling you a subscription on the promise that the paid tier draws pictures is describing something that does not exist.
Reading images is a different capability, and Claude has it. Upload a screenshot, a photo or a chart and it will describe what is there, pull the numbers out, or rebuild it as code. So the gap is output only, which is a narrower problem than most people assume when they first hit it.
| Task | Claude on its own | Evidence |
|---|---|---|
| Photo or illustration from a prompt | No | Anthropic help center |
| Diagram, chart, interactive visual | Yes, drawn in HTML and SVG | Anthropic help center |
| Read and analyze an image you upload | Yes | Anthropic help center |
| Call an outside image model | Yes, once connected | Connector documentation |
The two visuals Claude draws without any setup
Claude builds two kinds of visual with nothing installed: custom visuals, which appear inline in the chat and disappear when the conversation moves on, and artifacts, which are saved and shareable from the start. Both are drawn in code rather than generated as pixels, so they suit diagrams and charts, not photographs.
Custom visuals in chat
Ask directly and you get one. Phrases that work include “draw this as a diagram”, “show me how this changes over time” and “chart this data”. Anthropic’s documentation, updated April 22, 2026, lists custom visuals as a beta feature available to all Claude users on web and desktop. Opus handles complex visualization best.
Two limits catch people out. Custom visuals do not render on the iOS or Android apps, so a visual you made on a laptop will not display on your phone. And they are ephemeral by default. To keep one, copy it as an image, download it as an .svg or .html file, or save it as an artifact.
Artifacts, and the setting that has to be on
An artifact is the version you keep. Artifacts work on Free, Pro, Max, Team and Enterprise plans, and Claude can make one out of a diagram, a flowchart, a single-page website or a small interactive tool. Everything you make is saved to the Artifacts tab in the sidebar.
Here is the part worth writing down. Artifacts need “Cloud code execution and file creation” switched on in Settings, under Capabilities. On Team and Enterprise plans that toggle lives in organization settings. If artifacts are not appearing for you, check that before you conclude the feature is broken or that your plan is missing something.
Paid plans also get templates. Claude Design, available as a research preview on Pro, Max, Team and Enterprise, produces visuals, mockups, prototypes and landing pages on a canvas you can edit, and exports as a .zip, PDF, PPTX or standalone HTML. It is a design tool, not an image generator. It will build you a layout. It will not hand you a photograph.
How to generate images with Claude using a connector
This is the method people mean when they ask how to add image generation to Claude. You connect Claude to a remote server that runs an image model, and from then on you ask for pictures in the chat like anything else. The connection uses the Model Context Protocol, an open standard Anthropic created for linking AI applications to outside tools.
The setup steps, by plan
On Pro and Max, the route is short. Anthropic’s custom connector documentation, updated August 11, 2026, gives these steps:
- Go to Customize, then Connectors.
- Click the plus button, then “Add custom connector”.
- Paste the server’s remote MCP URL.
- Open Advanced settings if the server needs an OAuth client ID and secret.
- Click Add, then enable it per conversation from the plus button in the chat.
Team and Enterprise work differently. Only an Owner can add a connector to the organization, from Organization settings. Members then connect to it individually. That is a deliberate design: it keeps Claude’s access scoped to what each person already has rights to.
Connectors are available on Free, Pro, Max, Team and Enterprise, so the free tier is not locked out of this. It is capped instead.
Free users are limited to one custom connector.
Anthropic custom connector documentation, August 2026
The network rule that breaks most setups
This is the detail almost no guide mentions, and it explains a large share of failed connections. When you add a custom connector, Claude reaches your MCP server from Anthropic’s cloud infrastructure, not from your own computer. That holds across every client, including the desktop app and Cowork.
So your server has to be reachable from the public internet. A server behind a VPN, sitting on a corporate network, or blocked by a firewall will not connect, even though you can open it fine in your own browser. That mismatch is exactly what makes the failure confusing. The fix is to allowlist Anthropic’s IP ranges, or to host the server somewhere publicly reachable.
One more thing to know before you connect anything. Custom connectors link Claude to services Anthropic has not verified. Anthropic’s own documentation warns that a malicious server can carry hidden instructions designed to make Claude act against you, and advises connecting only to servers you trust and reviewing the permissions each one asks for. Treat an unknown image server the way you would treat an unknown browser extension.
Installing a desktop extension in Claude Desktop
Desktop extensions are the one-click version of the same idea. Instead of editing a JSON config file by hand, you install a packaged local MCP server from a directory inside Claude Desktop. Go to Settings, then Extensions, click “Browse extensions”, pick an Anthropic-reviewed tool, and enter any API key it asks for through the interface.
You probably do not need to install Node.js
Most setup guides open by telling you to install Node.js. Check that against the source. Anthropic’s documentation states that Claude Desktop includes a built-in Node.js environment, so installing Node is not required. Desktop extensions support Node.js, Python and binary MCP servers.
That correction saves a genuinely annoying step, and it is the kind of detail that gets copied between guides for years after it stops being true. API keys you enter are encrypted using your operating system’s secure storage, which means Keychain on macOS, Credential Manager on Windows, and your distribution’s keychain manager on Linux.
Confirming the server actually connected
Click the plus button at the bottom of the chat box, then select Connectors. That shows your connected MCP servers and the tools each one exposes. If an extension looks installed but no tools appear, restart Claude Desktop to refresh the extension registry, then check the extension’s settings for a missing required field.
Worth knowing: local MCP servers configured through the desktop config file are a separate mechanism from remote connectors, and they are not available in Cowork or on claude.ai. If a tutorial tells you to edit that file and you are working in the browser, the tutorial is for a different surface.
On Team and Enterprise plans, extensions may be restricted. Owners can turn public desktop extensions off entirely, or publish an allowlist so only approved tools install. If the directory looks empty or an install is blocked on a work machine, that is policy, not a bug, and your administrator controls it.
How to generate images with Claude Code
Claude Code connects to MCP servers from the terminal, which makes it the best route for batch work and for anything you want scripted. You add a server with one command, authenticate if it needs it, and then ask for images in plain language. Connectors you already added on claude.ai show up automatically.
Adding and checking the server
The Claude Code MCP documentation gives the syntax. A remote HTTP server is the recommended shape:
claude mcp add --transport http myimages https://mcp.example.com/mcp
claude mcp list
claude mcp get myimagesRunning the list command shows a health status beside each server: connected, needs authentication, or failed to connect. If it needs authentication, run the /mcp command inside a session and follow the browser sign-in. Servers are added at local scope by default, meaning only that project sees them. Add the user scope flag to make one available across all your projects.
Where the image file actually lands
This is the mechanic that decides whether you end up with a usable file. When an MCP tool returns a PNG, JPEG, GIF or WebP image, Claude sees a copy inline, which may be scaled down to fit the model’s size limits. Claude Code also saves the original bytes to a file in the session’s tool-results directory and hands Claude the path, so it can crop or convert the full-resolution version afterwards.
Two conditions apply, and both are easy to trip. Saving images to a file needs Claude Code version 2.1.283 or later. And if you have turned off session persistence, no image file is written at all and Claude only receives the scaled inline copy. If your generated images keep vanishing, that is where to look first. Once you have the file, compress it to WebP before it goes on a web page, because generated images tend to arrive far heavier than a page needs.
Use Claude to direct a different image tool
The sixth method needs no setup, works on every plan including free, and is the one most people should try first. You describe what you want, Claude writes the prompt, and you paste that prompt into whichever image generator you already pay for. Claude is good at the part that is hard, which is turning a vague idea into specific instructions.
Why connecting something is usually the wrong first move
Connecting an image model has real costs. You are trusting an unverified third-party server, you are managing an API key, and you still pay the image model for every generation. For someone who needs a handful of images a month, that is a lot of machinery for a small job.
Output quality also does not come from Claude. It comes from whichever model you connected. Claude adds value by directing, not by rendering, which means you get most of the benefit without any of the setup. The same pattern shows up elsewhere in AI work: the useful skill is usually briefing the tool well, which is the approach that makes AI-drafted speeches sound like the person giving them.
What a usable handoff prompt contains
Ask Claude for subject, setting, composition, camera angle, lens, lighting, depth of field, color palette and mood, then tell it which generator you are pasting into so it matches that tool’s conventions. Ask for three variations rather than one. Comparing options is faster than guessing at a single prompt and re-rolling it.
Keep the thread open. The real advantage is iteration: “same shot, warmer light, lose the gradient” works in a conversation, where it would mean rewriting the whole prompt by hand anywhere else.
A worked example. Instead of asking for “a photo of a coffee shop”, tell Claude the image is a hero banner for a roastery site, shot from counter height, shallow depth of field, warm morning light from the left, muted browns, no people in frame. Claude turns that into generator-ready wording, and you compare three versions before you spend a single credit.
Can you connect ChatGPT or Gemini to Claude for images?
Not the apps themselves. ChatGPT and the Gemini app are consumer products, and neither exposes its image generator to outside software. What you can connect is the API behind them. OpenAI’s Image API and Google’s Gemini API both generate images, and both reach Claude through an MCP server pointed at your own API key.
Why the apps cannot be wired together
People picture a switch that makes Claude ask ChatGPT for a picture. No such switch exists, in either direction. Both companies built MCP support so their assistant can call outside tools, not so outside tools can call their assistant. ChatGPT connects to your apps the same way Claude does, which makes the two of them peers, not components of each other.
Your subscription does not travel either. A ChatGPT Plus or Google AI Pro plan covers you using that company’s app. It grants nothing to Claude. The API is billed separately from the consumer subscription, so going this route means paying twice unless you drop the other subscription.
The image models you can actually reach
Both APIs are documented and open to anyone with a key. OpenAI’s image generation guide covers the generations and edits endpoints, and notes that GPT Image models may require organization verification in your developer console first. Google’s Nano Banana documentation covers the Gemini side.
| Provider | Model ID | Documented as |
|---|---|---|
| OpenAI | gpt-image-2.5-flare | Fast, high-quality generation |
| OpenAI | gpt-image-2.5-sunburst | Optimized for editing precision |
| gemini-3-pro-image | Nano Banana Pro, complex visual tasks | |
| gemini-3.1-flash-image | Nano Banana 2, balances speed and 4K | |
| gemini-3.1-flash-lite-image | Nano Banana 2 Lite, fastest and cheapest | |
| gemini-2.5-flash-image | Original Nano Banana, now legacy |
One naming note, because it confuses people. Nano Banana is not a separate product you sign up for. It is Google’s nickname for its Gemini image models, so reaching Nano Banana means calling the Gemini API with one of the model IDs above. Google lists the original gemini-2.5-flash-image as legacy and points new work at Nano Banana 2 Lite.
Neither company ships an official connector
Here is the honest state of it. Searching the MCP registries turns up many servers wrapping OpenAI and Gemini image generation, and every one found in October 2026 was community-built. Neither OpenAI nor Google publishes an official MCP server for image generation, so there is no vendor-supported option to recommend.
That leaves two defensible routes. Run a community server and accept the risk already described above, after reading its source and checking what it does with your API key. Or skip the server entirely: in Claude Code, hand Claude your key as an environment variable and ask it to write a short script that calls the endpoint directly. The second route has no third party in it at all, which is why it is the one to prefer for anything sensitive.
Which way to generate images with Claude fits you
Pick by what you need the picture for. Diagrams and charts are free and instant. Real photographs need an outside model, which means either a connector or a copy and paste. Nothing here makes Claude itself draw a photo, so the choice is really about how much setup you want between you and the model that can.
One rule of thumb cuts through the table below. Count how many images you need in a month. Under about ten, the clipboard wins and setup is wasted effort. Over a hundred, scripting against an API pays for itself quickly. The middle is where a connector earns its place, and it is a narrower band than most guides suggest.
| Method | Setup | What you get | Best for |
|---|---|---|---|
| Custom visuals in chat | None | Diagrams, charts, comparisons | Explaining something once |
| Artifacts | One capability toggle | Saved, shareable visuals | Work you reuse |
| Claude Design | Paid plan, preview | Mockups, layouts, prototypes | Design work, not photos |
| Custom connector | Server URL, often OAuth | Real generated images in chat | Regular image work |
| Desktop extension | One click plus an API key | Real generated images, locally | Claude Desktop users |
| Claude Code plus MCP | One command | Image files on disk | Batches and scripting |
| Claude Code plus an API key | A key and a short script | OpenAI or Gemini images, no third party | Anything sensitive |
| Claude writes the prompt | None | A prompt you paste elsewhere | Almost everyone starting out |
Why can’t Claude generate images?
Because there is no image model attached to it. Claude is built to understand and produce text, to read images through a vision encoder, and to reason. Text-to-image generation is a separate model class that Anthropic has not shipped. Anthropic has published no roadmap saying it will, and anyone telling you a date is guessing.
Running image models is its own infrastructure problem, separate from serving a language model, which is a recurring theme in how AI infrastructure costs actually behave. Whether that is the reason or simply a product choice, Anthropic has not said, and this guide is not going to invent a motive.
Frequently asked questions
Can the Claude model generate images?
No. No version of Claude generates photos or illustrations from a text prompt, and that applies to every plan from Free through Enterprise. Anthropic states this in its own help center. Claude does produce visual output in the form of diagrams, charts and interactive visuals written in HTML and SVG, and it can read and analyze images you upload to it.
Is Claude free to generate images?
Claude’s own visual output costs nothing. Custom visuals are in beta for all users on web and desktop, and artifacts are available on the Free plan. Connecting an outside image model is also possible on Free, but free accounts are limited to one custom connector, and the image model itself usually charges per generation. So the connector is free, the pictures may not be.
How do I add image generation to Claude?
Add a remote MCP server that runs an image model as a custom connector. On Pro or Max, go to Customize, then Connectors, click the plus button, choose “Add custom connector”, paste the server URL, and add OAuth credentials under Advanced settings if the server requires them. On Team and Enterprise, an Owner adds it at the organization level first, then members connect individually.
Why does my image connector say failed to connect?
Most often because the server is not reachable from the public internet. Claude connects to your MCP server from Anthropic’s cloud, not from your computer, so a server behind a VPN or a corporate firewall fails even though it opens fine in your browser. Allowlist Anthropic’s IP ranges, or host the server somewhere publicly reachable. Authentication errors are the other common cause.
How do I generate my own AI image?
Use a tool built for it, and let Claude write the prompt. Describe the picture you want, ask Claude for subject, setting, composition, lighting, lens and mood, request three variations, then paste the result into your image generator. This needs no setup, works on any plan, and plays to the thing Claude is genuinely good at, which is turning a loose brief into specific instructions.
Is there a 100% free AI image generator?
Several services offer free tiers, and some open-weight models are free to run on your own hardware. Claims that any hosted service is permanently and completely free deserve scepticism, because generation costs the provider real compute every time. Free tiers in this category change pricing often, so check the provider’s current terms rather than a blog post, including this one.
Does Claude Design generate images?
Not in the sense people usually mean. Claude Design is a research preview on Pro, Max, Team and Enterprise that produces designs, mockups, prototypes and landing pages on an editable canvas, and exports as a .zip, PDF, PPTX or standalone HTML. It builds layouts and interfaces, not photographs. If you need a product photo, Claude Design will not give you one.
Can Claude use ChatGPT to generate images?
Not the ChatGPT app, no. It does not expose its image generator to other software, and a ChatGPT subscription grants Claude nothing, because the API is billed separately from the consumer plan. You can reach OpenAI’s Image API instead, using your own API key and models such as gpt-image-2.5-flare for speed or gpt-image-2.5-sunburst for editing. OpenAI notes that GPT Image models may need organization verification in your developer console first.
Can Claude use Nano Banana?
Yes, through the Gemini API, and not through the Gemini app. Nano Banana is Google’s nickname for its Gemini image models rather than a separate product, so you call the Gemini API with a model ID like gemini-3.1-flash-image for Nano Banana 2 or gemini-3-pro-image for Nano Banana Pro. Google marks the original gemini-2.5-flash-image as legacy and points new work at Nano Banana 2 Lite instead.
Is there an official OpenAI or Google MCP server for images?
Not as of October 2026. The MCP registries list many servers wrapping both providers’ image endpoints, and all of them are community projects rather than vendor-published software. That matters for trust, because a server you connect can read whatever you send it. Writing a short script against the API in Claude Code avoids handing your key to a third party at all.
Can Claude edit an image I upload?
It can read one and describe it, but it cannot repaint it. Image understanding and image generation are separate capabilities, and Claude has the first without the second. It can rebuild what it sees as code, so asking it to recreate a chart you uploaded as an editable SVG works well. Asking it to retouch a photograph, remove an object or change a background does not.
Why can’t I see the visual Claude made on my phone?
Custom visuals render on Claude web and desktop only. They do not display in the iOS or Android apps. If you shared a chat containing one, the recipient also needs to be logged in and on web or desktop to see it. Saving the visual as an artifact first is the way to make it travel, since artifacts can be viewed in the mobile Artifacts tab.
What file formats can I download a Claude visual as?
A custom visual downloads as an .svg or .html file, or you can copy it as a static image. Designs made from the Claude Design template export as a .zip, PDF, PPTX or standalone HTML. SVG stays sharp at any size, which makes it the better pick for diagrams. These are text-based formats, much like Markdown files, so you can open and edit them in any text editor.
Getting an actual picture, finally
That failed connector you started with was probably not your fault, and probably not broken software either. It was most likely a server Anthropic’s cloud could not reach, or a capability toggle sitting in the off position. Both take two minutes to check once you know they exist, and neither appears in the guide that told you to install the thing.
The wider point is duller and more useful. Learning how to generate images with Claude mostly means learning what Claude is for. It is a very good director and not a camera. If you need diagrams, it already makes them and you do not have to install anything. If you need photographs, something else has to render them, and your only real decision is whether that handoff happens through a connector or through your clipboard. For most people, most of the time, the clipboard is enough.
Which leaves a question worth sitting with before you start wiring things together. How many of the images you think you need are really photographs, and how many are diagrams you have been paying an image model to draw badly?
