You connect image generation to Claude Desktop, Cursor, and ChatGPT by pointing each client at the same MCP server and authenticating with one API key. MCP is a tool protocol, not a model: a server advertises functions, the agent calls them, the server runs them. For a still image that call is synchronous, and the response body contains a hosted HTTPS URL. You do not write a wrapper per client, and you do not stand up a bucket to hold the file. PicX's REST API lives at https://api.picxstudio.com. Keys start with pxsk_. Auth is Authorization: Bearer $PICX_API_KEY. That same key covers all three surfaces.
A server exposes tools. The agent calls them
A coding agent is a language model plus a harness. The model emits tokens. The harness edits files, runs commands, and invokes tools. Image models live outside that process. Without a generation tool, "make a PNG of a ceramic mug on wet slate" becomes a paragraph, SVG, or a refusal. Claude Desktop and Cursor have no native image tool. They speak tools. If none is connected, there is nothing to call.
MCP, the Model Context Protocol, is how you attach one. The server publishes a list: name, description, argument schema. The client shows that list to the model. When the model wants an image, it emits a tool call with a prompt and optional size or aspect ratio. The client forwards the call with your auth header. The server hits the generation API and returns a result the model can read, including a URL the rest of the toolchain can fetch.
The clients differ in where you paste a URL. They do not differ in the HTTP that follows. One remote MCP server, Streamable HTTP, stateless. Every request carries the key. There is no OAuth handshake and no session cookie. If a client tries OAuth anyway, the connector looks saved and no tools appear.
PicX's MCP server exposes 19 tools covering generate, edit, video, assets, templates, and account. You will use generate first. Image generation is synchronous: the tool waits, the body includes a hosted image URL, there is no poll. Video and batch jobs are asynchronous and delivered by webhook. Teach the agent the wrong contract and it either spins on work that already finished or hangs on a render that was never going to return inline.
Output files live on PicX's own CDN. You do not bring object storage. You do not configure a second host. The URL in the tool result is the file.
One key covers Claude Desktop, Cursor, and ChatGPT
Create one key. Send it as a Bearer token. Point three clients at the same MCP server. That is the integration.
The key is the same credential the REST API, the CLI, and the official Python and JavaScript SDKs (picx-ai on npm and PyPI) already use. Roughly 33 image and video models sit behind it. Billing is in credits per generation, never a currency amount on the meter. Set a hard spend cap on the key before an agent can retry a prompt in a loop.
Scopes on the key gate the tools the same way they gate REST endpoints. A client can list a generate tool and still fail if the key was minted without image-generation scope. Grant that scope for this walkthrough. Add edit and video when you want those calls. Do not mint a key per client unless you want separate caps or revocation. The protocol does not require it.
Connection docs live at https://picxstudio.com/developers/mcp. ChatGPT connects to the same MCP server; setup is on https://picxstudio.com/developers/mcp. Agent Skills are at /skills. If you want a coding agent to do the wiring, https://picxstudio.com/developers/agent-setup/prompt.md is an executable checklist. Live `llms.txt` and `llms-full.txt` are the machine-readable reference. None of those replace the header. They stop you from inventing parameter names.
Mint the key before you touch a client
Sign in to the developer console and create an API key. The full pxsk_ value is shown once. Store it in a secret manager or an environment variable. Lose it and you revoke, then mint another. Keep it out of git and out of any browser bundle. Cursor's project mcp.json is the usual leak: the file lives in the repo, so people commit it. Claude Desktop and ChatGPT keep the header inside the app — better for secrets, worse for a shared team config.
Connecting costs nothing. Generation tools deduct credits when they actually run. A verification prompt is a real charge. Set the spend cap first.
The MCP host is not the REST host. REST calls go to https://api.picxstudio.com. The MCP server is a separate origin, documented on /mcp, mounted at that host's root — no /mcp or /sse path suffix. Mixing the two is the most common 404 in this setup.
Claude Desktop: a connector, not a config file
Claude Desktop adds remote MCP servers from Settings. You do not edit a JSON file for this path.
Open Settings → Connectors. Add a custom connector. Name it something you will recognize. Paste the PicX MCP URL from the /mcp page — the root, not the REST base, not a path the server does not serve. Set the connector's Authentication control to none. PicX authenticates with a header, not OAuth. Add a header named Authorization with value Bearer plus the pxsk_ key, one space after Bearer. Save. Open a new conversation. If the 19 tools do not show, restart the app. Versions disagree on whether a new chat is enough.
The failure mode that looks like success is the OAuth dropdown. Leave it on OAuth and Claude attempts a handshake the server does not implement. The connector exists. The tool list is empty. The header is the authentication.
Then ask, in a fresh chat: generate an image of a red apple on a white background. You want a tool call and a URL. If Claude offers a stock-photo connector or says it cannot generate images, either the header never saved or the prompt read as "find a photo." Generation intent is generate, create, draw. Search intent is find, look up.
Cursor: project or global mcp.json
Cursor reads MCP servers from a JSON file. Project-scoped: .cursor/mcp.json at the repo root. Global: ~/.cursor/mcp.json. Use the project file when the key belongs to that codebase. Use the global file when one key covers every project you open.
The shape is a map named mcpServers. Each entry needs a url and a headers object. The header is the same Bearer line. Cursor does not expand environment variables inside that file, so a literal key is the working form:
{
"mcpServers": {
"picx": {
"url": "<mcp-server-url>",
"headers": {
"Authorization": "Bearer <pxsk_key>"
}
}
}
}Fill url from the /mcp page. Treat the file as a secret; if it sits in the repo, gitignore it. Merge the PicX entry into an existing mcpServers object. Replacing the file deletes every other server you had configured.
Save. Cursor usually picks up the change without a restart. If the tools do not appear, reopen the MCP panel or restart. Then run the same apple prompt. Success is a visible tool call, a credit debit, and a CDN URL you can open. A description with no tool call means the server never attached. Check the host and the header spelling before you rewrite the prompt.
The CLI can write this config (picx mcp install, then picx mcp doctor) if you would rather not touch the file.
ChatGPT: the same server, a different UI
ChatGPT attaches the same MCP server through its connector UI, not through a second API and not through a plugin manifest. The facts you must supply are the MCP URL and the Bearer header with the pxsk_ key. The labels around those fields change as ChatGPT ships connector updates. https://picxstudio.com/developers/mcp tracks the current setup steps. Do not copy a JSON blob from a write-up of a different MCP server. ChatGPT does not publish a stable on-disk config the way Cursor does.
This MCP server is stateless HTTP. Every request carries the key. There is no OAuth client id to mint. If a given ChatGPT build only offers OAuth and has no custom-header field, the handshake will not complete. Use the header path, or generate from Claude Desktop or Cursor until the UI exposes one.
After it connects, the model sees the same 19 tools. A generate call still returns a hosted URL. ChatGPT may also expose a native image tool. If you wanted the PicX meter and the PicX CDN, say so: generate with the connected PicX server. Otherwise the model may pick the built-in generator and you will wonder why the spend cap never moved.
Verify once, then use the URL
Do this once per client you actually use. A connector saved without a header looks identical to a working one until you ask.
Approve the tool call if the client prompts. Confirm three things: a PicX tool name, a credit charge, and an HTTPS URL that loads in a browser.
Follow-up edits take that URL as input. If you need to send a local file, upload it with the asset tool first. Local paths and data URIs are the wrong type. Video is the other contract: the generate-video tool returns an id, and completion arrives by webhook. Do not poll the image path. It is already done when the tool returns.
If you would rather skip MCP, POST to https://api.picxstudio.com with the same key. The playground, the CLI and the picx-ai SDKs all hit that API. MCP is how Claude Desktop, Cursor, and ChatGPT already invoke tools. HTTP is how a backend already invokes APIs. Same pixels. Same CDN URL.
What you give up
The pixels leave your machine and live on a hosted CDN. If that is unacceptable, run a local model and operate storage yourself. You pay in credits, not idle GPU time. Spend caps bound a runaway agent. They do not make a bad prompt cheap.
MCP is not automatic. A missing header, an OAuth dropdown, a prompt that asked to find a photo — all of those produce prose again. Check the tool list. Generate once on purpose. SVG is still the right output for icons and diagrams. Do not send every "draw" through an image model.
FAQ
Do I need a different API key for Claude Desktop, Cursor, and ChatGPT?
No. One pxsk_ key, sent as Authorization: Bearer $PICX_API_KEY, authenticates the MCP server and the REST API at https://api.picxstudio.com. Use separate keys only when you want separate spend caps or revocation.
Why did the agent describe an image instead of generating one after I connected?
The client either did not attach the MCP server, or the prompt read as a request to find an existing photo. Confirm the 14 PicX tools appear in the client's tool list, then ask it to generate a new image.
Does a simple image generation require polling?
No. Image generation is synchronous: the generate call returns a hosted image URL in the response body. Video and batch jobs are asynchronous and delivered by webhook.
Where are the generated files stored?
On PicX's own CDN. You do not bring or configure a separate file-hosting service. The tool result includes an HTTPS URL you can embed or fetch directly.
