Image Generation MCP Server: Hosted, No Install

An image generation MCP server lets your AI assistant make images for you as a tool call, and GenBatch runs a hosted one you add by pasting a single URL. No install, no separate model API key, no server to keep alive. You ask Claude, ChatGPT, Gemini CLI, Cursor, or another MCP-capable client for images in plain language, and they come back in the conversation, charged to your own GenBatch credits at one credit each.

That last part is the difference. Most image generation MCP servers you find are code you have to clone, run, and wire to your own OpenAI or Replicate account. GenBatch is the opposite: it is already running, and your account is the only thing you connect.

What is an image generation MCP server?

It is a server that exposes image generation as a callable tool over the Model Context Protocol (MCP), the open standard AI clients use to talk to external tools. Instead of you opening a separate app, pasting a prompt, waiting, and downloading, your assistant calls the tool directly and hands you the result.

MCP is the same mechanism that lets Claude read a file or query a database. An image server just adds a generate_image tool to that list. Once it is connected, "make me 5 square thumbnails of a matte black water bottle" becomes a single tool call your assistant runs on its own.

The catch with most of them is setup. Search for one and you land on GitHub repos that wrap Flux, DALL-E, or gpt-image and expect you to run the process locally and bring your own paid API key. That works, but it is infrastructure you now own. GenBatch skips that entirely.

Why use a hosted MCP server instead of a self-hosted one?

Because a hosted server removes the two jobs you would otherwise inherit: running the process and supplying a model key. GenBatch is already live, stateless, and reachable at one URL, so you connect your account and start generating.

Here is the practical split between the two approaches.

We think the self-hosted route makes sense if you want deep control over a specific open model and enjoy running infrastructure. For everyone else, a hosted server is less to break. There is no process that dies overnight, and because the GenBatch server is stateless, it keeps working across restarts and multiple replicas without any session juggling on your side.

How do you connect GenBatch as an MCP server?

Three steps: choose an auth path, add the remote server, then ask for images. The whole thing takes about a minute, and you never leave your assistant after that.

Claude supports a login-based connector flow: paste the URL, sign in to GenBatch, and approve access. In ChatGPT, enable developer mode in Settings -> Apps & Connectors -> Advanced settings, then create a connector with the GenBatch MCP endpoint. In Gemini CLI, add the HTTP MCP server with gemini mcp add --transport http and an Authorization: Bearer gbm_your_key header. The bearer-token flow also works in Cursor, MCP Inspector, and other MCP clients that accept custom authorization headers. If you want the full walkthrough for Claude specifically, see our guide on generating images in Claude.

What can the GenBatch MCP tools do?

The core image loop is still simple: start a job, check on it, and read your balance. These tools map one-to-one onto how batch generation already works inside GenBatch, so nothing behaves differently just because you called it from Claude.

• generate_image submits a job. For one prompt, set prompt and optionally number_of_images (1 to 10), aspect_ratio (16:9, 9:16, 1:1, 4:3, or 3:4), and image_model (gpt_image_2 by default, or nano_banana_2). For multi-prompt work, pass items[] with up to 20 prompts and 50 total images in one job.
• check_job takes that job_id and reports status. While the job runs it says not ready. When it finishes, it returns a flat files URL list, a gallery_url when available, and an images grouping for multi-prompt jobs.
• get_credits returns your current spendable balance, so your assistant can warn you before you run low.
• get_balance, list_credit_packages, and buy_credits_from_balance let Claude inspect wallet balance and buy a credit package from existing wallet funds.
• text_to_video and frames_to_video submit video jobs. A URL returned by check_job.files from an image job can be passed into frames_to_video as start_frame_url.

Because these tools call the same job pipeline as the GenBatch app, a batch you kick off from Cursor is the same batch you would see in your dashboard. Bulk is the point here. Asking for 10 copies of one prompt is one call, and asking for several different prompts can also stay grouped as one batch through items[].

How much does an image generation MCP server cost?

With GenBatch, one credit per image. Ask for 4 images and it costs 4 credits. There is no subscription and no per-seat fee, and your credit balance is the real spending cap.

Billing is pay-on-success. Credits are checked upfront but only deducted after a generation actually succeeds, so a failed request does not cost you anything. If your balance is too low, the request is rejected with a clear message and nothing is charged. That is worth calling out, because a self-hosted server bills you for the model API call whether the result is good or not.

Is an image generation MCP server secure?

It can be, and GenBatch is built so a leaked key is a small, reversible problem. The access key is opaque, meaning it carries no readable claims like a JWT does. It is a gbm_ prefix followed by 64 random hex characters, and only its SHA-256 hash is stored on our side. We literally cannot show it to you again after creation.

A few more guardrails worth knowing:

• Revocable and capped. You can hold up to 10 active keys and revoke any of them instantly at /connect. Revoking kills both the MCP and REST access for that key at once.
• Scoped to you. A key only ever acts on its owner's account and credits. It cannot touch anyone else's.
• Ownership checks on results. When check_job looks up a job, it verifies you own that job. Passing an id you do not own returns "no job found" with nothing leaked about whether that job exists.
• Abuse backstops. Sliding-window rate limits (200 generate calls per 10 minutes, 120 status checks per minute, 60 balance reads per 10 minutes, and 3 wallet purchases per 10 minutes) sit on top of credits as a runaway guard. In normal use your credit balance is the limit you notice, not these.

Can I use it from my own code instead?

Yes, and it is the same key. If you would rather call GenBatch from a script or backend than through an assistant, the REST API takes the exact gbm_ value you already created. Submit a job, then poll for the result.

curl -X POST https://genbatch.com/api/queue/submit \
  -H "Authorization: Bearer gbm_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "create-image",
    "requestData": {
      "prompt": "a red fox in snow",
      "number_of_images": 1,
      "aspect_ratio": "1:1",
      "image_model": "gpt_image_2"
    }
  }'

The response includes a job id. Poll GET /api/results/<id> until the status is completed, and the image URLs come back in result_data.files[]. For the full endpoint list, parameters, and error codes, see the GenBatch image generation API reference. And if bulk is your main use case, our guide on generating 100 AI images at once covers the batch workflow end to end.

Frequently asked questions

What is an image generation MCP server?

It is a server that exposes image generation as a tool your AI assistant can call over the Model Context Protocol. Once connected, you ask Claude, ChatGPT, Gemini CLI, Cursor, or another MCP client for images in plain language and it calls the tool for you. GenBatch runs a hosted one, so there is nothing to install.

Do I have to install or run the GenBatch MCP server myself?

No. GenBatch is a hosted remote MCP server. You add it by pasting the endpoint URL into your MCP client. Claude can use the login connector flow; clients without that flow can use a GenBatch access key. Most other image MCP servers are self-hosted, which means you run them locally and supply your own model API key.

How much does it cost to generate an image?

One credit per image. Asking for 4 images costs 4 credits. It is pay per use with no subscription, and billing is pay-on-success, so a failed request costs nothing. If your balance is too low the request is rejected and nothing is charged.

Which MCP clients work with GenBatch?

Any client that supports remote MCP servers, including Claude, ChatGPT developer-mode connectors, Gemini CLI, Cursor, and the MCP Inspector. Claude and ChatGPT can use connector flows. Gemini CLI, Cursor, and manual clients can use the same endpoint with a GenBatch bearer key.

Is one access key enough for both MCP and the REST API?

Yes, for manual setup. One gbm_ key works for the hosted MCP server and the REST API. Claude can also use the no-key login connector flow. You can hold up to 10 active keys per account and revoke any of them instantly at /connect.

Which image models can the MCP server use?

You can pick gpt_image_2 (the default) or nano_banana_2, set the aspect ratio (16:9, 9:16, 1:1, 4:3, or 3:4), generate 1 to 10 copies of one prompt, or use items[] for up to 20 prompts and 50 images total in one call.