MCP server
Connect Claude Desktop, Cursor or a local agent to XPlor3, then ask the agent for images and video from the conversation.
XPlor3 runs an official Model Context Protocol server. Once it is connected, an AI agent in an MCP client can browse the model catalogue, start a generation, wait for the result and hand you a download link, all from an ordinary conversation.
Address
https://mcp.xplor3.space/mcpThe transport is HTTP (streamable HTTP), with the API key in the Authorization header. No key yet? Make one first in Credentials.
Connecting a client
Paste this block into your client's MCP configuration file, then replace xpk_your_key with your real key. The studio's Credentials page shows the same block, ready to copy.
{
"mcpServers": {
"xplor3": {
"type": "http",
"url": "https://mcp.xplor3.space/mcp",
"headers": {
"Authorization": "Bearer xpk_your_key"
}
}
}
}Open Settings > Developer > Edit Config, paste the block above into claude_desktop_config.json, save, then restart Claude Desktop. The xplor3 server appears among the conversation's tools.
Open Settings > MCP and add a new server, or paste the block above into .cursor/mcp.json in your project (or ~/.cursor/mcp.json for every project).
Any client that supports MCP servers over HTTP with custom headers can use the same address and header. Clients that only support stdio need an HTTP-to-stdio bridge from that client itself.
Available tools
| Tool | What it does |
|---|---|
whoami | The account that owns the key and whether it may spend credits |
get_balance | The current credit balance |
list_models | Every model the account can run, with its kind, maker and starting cost |
describe_model | One model's fields, defaults and variants, so a valid input can be built for generate |
generate | Starts a generation and returns the job id, the cost and the balance left. Takes an optional webhookUrl |
wait_for_generation | Waits on the server until the job ends, then returns it. One call, no polling |
get_generation | One job's status and the ids of its result files |
cancel_generation | Cancels a job that is still queued and returns its credits |
list_generations | The account's generations, newest first |
get_asset_links | Short-lived download links for result files |
upload_reference | Uploads a reference image, video or audio as base64 |
list_projects | The account's projects |
create_project | Creates a new project |
add_to_project | Files results into a project |
The usual flow an agent follows: list_models, then describe_model for the chosen model, then generate, then wait_for_generation, then get_asset_links.
wait_for_generation waits up to 50 seconds by default, and up to 120 with timeoutSeconds. An image usually ends within one call. For a video the answer can be finished: false with the latest status and progress, and the agent simply calls it once more.
Every model's input fields, with the values they accept and the rules per tier or mode, are in the Model reference. describe_model returns the same data as JSON.
Example request
Once connected, just ask in plain language:
Make three variations of a product photo of white running shoes on an athletics track at sunset, with Nano Banana 2 at 1K, then file the results into the Ads project.
The agent checks the balance and the cost, runs the generation, waits for the result, then gives you the download links.
Good to know
- Credits are taken when
generateis called, not when it finishes, at the same rates as the studio. The reply carries the cost and the balance left. - Cancel only while queued. A job already sent to the model cannot be cancelled (
JOB_ALREADY_SENT) and its result is still delivered. - Image and video models. Speech models and Workflow are not available over MCP yet.
- The concurrency and request limits are the same as the REST API. See Limits and errors.