Midjourney still has no public API, so Claude cannot call it directly. Here are the three workarounds in use today, browser-driven MCP servers, resold API proxies and manual hand-off, with their real risks, setup commands and a supported alternative for reliable image generation.
You ask Claude for a moody editorial portrait "in Midjourney" and expect a finished image to land in the chat. It won't, because Midjourney has no public API for an assistant to call. That single fact decides everything else in this post. The honest answer to "can Claude use Midjourney?" is yes, through side doors that Midjourney does not officially support. Below you'll see which side doors exist, what each one lets Claude do, where they break, and what a supported alternative looks like when you need a setup that still works next month.
At the time of writing, no. Midjourney offers two official ways to make images: the web app at midjourney.com and its Discord bot. Both sit behind a paid subscription that starts around $10 per month. There is no public endpoint, no credential to generate, and no per-image price list. Most rivals work the opposite way. Flux, Imagen, GPT Image and Seedream all ship with ordinary developer APIs, which is why an assistant can connect to them in minutes.
Midjourney's terms of service prohibit using automated tools to access, interact with, or generate assets through the service. The current version is reported as effective May 27, 2026. Users report accounts being blocked for "accessing our services via automation", and one source notes that a banned account loses whatever GPU time was left on it.
💡 Read this before anything else: every route in this article relies on automation, and Midjourney's terms prohibit it. Treat each one as a convenience with a real chance of costing you your subscription, not as a supported integration.
Claude reaches outside tools through the Model Context Protocol (MCP). An MCP server publishes a list of tools, such as "generate an image" or "check job status", and Claude calls them mid-conversation. When a service has an API, the server is a thin wrapper around HTTP requests. Midjourney gives it nothing to wrap, so any "Midjourney MCP" has to do one of two things: drive a real browser session, or hand your prompt to a middleman who does.
Every working setup falls into one of three buckets. The table shows how they compare before we look at each one.
| Route | How it connects | Setup effort | Main risk |
|---|---|---|---|
| Browser-driven MCP server | Controls Chrome signed in to your account | Medium: Node 22+ and Chrome | Account flags, interface changes |
| Resold API proxy | A third party automates accounts and sells an "API" | Low: paste a credential | Terms, outages, trusting a reseller |
| Manual hand-off | Claude writes prompts, you paste them | None | Slow, but the safest of the three |
The most capable option listed in public MCP directories is a server by Navid Moazzez, published as @thenavidm/midjourney-mcp-cli. Its listing advertises 32 tools for generating images, following jobs until they finish, downloading the real files and building moodboards. Because Midjourney publishes no API, it works by driving a real Chrome window signed in as you. It needs Node 22 or newer plus Google Chrome.
The upside is that results come back as actual files, not screenshots of a web page. The downside is that it behaves like a very fast, very consistent user, and that is exactly the pattern anti-automation rules are written to catch.
A second family of servers talks to third-party "Midjourney API" services. AceDataCloud's MCPMidjourney is listed as connecting through an API rather than through a browser. The ImaginePro MCP server lists eight tools: text to image, multimodal generation, animating frame images, upscaling, variants, inpainting and task status. You authenticate with a credential passed as an environment variable and launch it with npx.
One detail from the ImaginePro listing deserves attention. It advertises Midjourney, Flux and Nano Banana together, so check which model actually renders each request before you judge the output.
Reports describe these services as automating consumer accounts through Discord or the web app and reselling the output. That moves the terms problem to someone else, but it also means a stranger sees every prompt you send and every account behind the curtain can be flagged.
The lowest-tech route needs no server at all. Ask Claude to write the prompt, paste it into Midjourney yourself, then come back with a description or a screenshot of the result and ask for a revision. You are the one pressing the button, so nothing is automated and the terms are not in play. It is slower, but for a few images a day it is hard to beat.
Even in the best case, Claude does not render anything. It is a prompt writer and job runner. Midjourney still produces the pixels on its own servers, and what changes is how much typing, waiting and filing you do by hand.
Claude is genuinely strong at the part people find tedious: turning a loose idea into a structured prompt and changing one variable at a time. Midjourney's parameters are short and easy to reason about: --ar for aspect ratio, --stylize for how far it drifts from your wording, --chaos for variety, --seed for repeatability, --no to exclude elements and --sref to borrow a style.
A request like "a ceramicist in her studio, window light, muted earth tones" can come back as:
documentary portrait of a ceramicist at her wheel, soft window light from the left,
85mm lens, shallow depth of field, clay-dusted apron, muted earth palette --ar 3:2 --stylize 250 --chaos 10
Ask Claude to keep a running log of the prompt, seed and parameters for every image you like. That log is what lets you reproduce a look three weeks later.
With a browser-driven or proxy server, Claude can go further than writing text:
With manual hand-off, Claude can still do the bookkeeping: file names, a prompt table and a short summary of what each round taught you about the look you are chasing.
Three weaknesses show up again and again, and none of them is fixable from your side.
This is the big one. Automation violates Midjourney's terms, so a flagged account can lose access and any unused GPU time. Using an account you can afford to lose, keeping batches small and never reselling output all reduce the odds. None of it removes the risk.
A browser-driven server clicks buttons on a page it does not control. When Midjourney redesigns part of its site, the server can return errors or do nothing until its maintainer ships a fix. Pin the version you tested and check the project's issue tracker before blaming your own setup.
Midjourney publishes no API price list. On a browser-driven setup, every image draws on your plan's fast GPU time. On a proxy, you pay the reseller's own per-image rate. Neither gives you the clean, predictable billing you would get from an official API.
| Problem | What you notice | Practical fix |
|---|---|---|
| Account flag | Jobs rejected or access removed | Low volume, one job at a time, a disposable account |
| Interface change | Server errors or silent no-ops | Pin the version, watch the issue tracker |
| Proxy outage | Timeouts, missing images | Keep a fallback model ready |
| Surprise cost | Fast hours drained early | Cap batch sizes in your instructions to Claude |
If you accept the risk, setup is short. These commands come from the server's own listing, so check its README for the exact sign-in step before you run anything.
You need Node 22 or newer and Google Chrome. Then one command registers the server:
claude mcp add midjourney -- npx -y @thenavidm/midjourney-mcp-cli@latest
Open claude_desktop_config.json and add the server under mcpServers:
{
"mcpServers": {
"midjourney": {
"command": "npx",
"args": ["-y", "@thenavidm/midjourney-mcp-cli@latest"]
}
}
}
Restart Claude Desktop, make sure the Chrome session the server drives is signed in to your Midjourney account, then try a small test: one prompt, one image, and ask Claude to report the job status. Only scale up once that works.
💡 Tip: start with a single job and watch it run. If the first test behaves oddly, a ten-image batch will behave worse.
If you want image generation that Claude can run all day without worrying about a ban, pick a model that ships an official API. You can draft the prompt with Claude Sonnet 5 or Claude Opus 4.7 and render it with a text-to-image model, all from one place.
| Model | Why try it |
|---|---|
| Seedream 4.5 | 2K or 4K output, up to 15 related images per run, 1 to 14 reference images |
| Flux 2 Pro | Up to 8 reference images, seed-based repeatability, up to 4 MP |
| GPT Image 2 | A strong pick when the image needs legible text |
| Imagen 4 Ultra | Google's top Imagen tier, a natural fit for photographic realism |
Reference images are the closest match to Midjourney's --sref. Upload two or three frames in the style you like, describe the new subject, and the model borrows the look.
-- parameters, which only Midjourney reads. Describe the subject, lens, light and palette in plain sentences.2K is the default and quick for drafts. Switch to 4K for the final render.match_input_image if you upload a reference.auto and raise the max images value, up to 15 related images in one run.| Setting | Best for | Tip |
|---|---|---|
size: 2K | Fast drafts | Use it while you test wording |
size: 4K | Final renders | Run it once the prompt is settled |
max_images up to 15 | Moodboards and series | Describe the shared scene once |
image_input 1 to 14 | Style matching | Pick references with one consistent look |
💡 Want repeatable results? Flux 2 Pro accepts a seed, so reusing the same seed with the same prompt recreates the same image.
Picasso IA also runs the integration Midjourney does not. It has a developer API at api.picassoia.com/v1 that follows the familiar create, poll and fetch pattern for asynchronous jobs, and an MCP connector you manage from your account's MCP connections page. At last check, the connector exposes four models, including PicassoIA Image and PicassoIA Image Editor Pro, and each account can run five predictions at once.
That is the difference that matters here. You connect it once, ask Claude for an image and it arrives, with no browser to babysit and no terms to worry about.
You now have the full picture. Midjourney through Claude is possible, but it is a workaround that can break or cost you an account at any time. A supported model gives you the same prompt-to-picture loop without that risk.
Take the portrait idea you were about to send to Midjourney and run it on Seedream 4.5 or Flux 2 Pro on Picasso IA. Change one detail, run it again and compare. Ten minutes of testing will show you which model fits your style, and you can keep Claude in the loop for prompts the whole way.
