Docs · Integrations
Hosted MCP servers: models, images and video in any MCP client
Three MCP servers put the catalogue in Claude Code, Cursor, VS Code and other MCP clients with your API key. Setup, tools and what a call costs.
TopxAI runs three Model Context Protocol servers. An MCP client that connects to them can ask any catalogue model, generate images and generate video with the same API key and the same prices as the API. The MCP page builds the setup lines for each client with your key filled in.
The three servers
| Server | Address | Tools | Fee per tool call that reaches a model |
|---|---|---|---|
| TopxAI Models | https://api.topxai.xyz/mcp/models |
list_models, ask_model, judge |
$0.001 |
| TopxAI Images | https://api.topxai.xyz/mcp/images |
generate_image |
$0.01 |
| TopxAI Videos | https://api.topxai.xyz/mcp/videos |
generate_video, video_status |
$0.01 |
ask_model sends a chat request to the model you name; judge sends a text question to Jev (image assets and Box questions stay on the HTTP endpoint). generate_image and generate_video are the image and video endpoints. list_models and video_status are free.
What a call costs
A tool call that reaches a model is billed exactly like the same API request: the model's price on your key's route, the same reservation and refund rules, one row in the usage log. The server's call fee is added to that request as its own line and is refunded with it if the request fails. Listing tools and polling a video cost nothing.
Setup
Authentication is your API key as a bearer token; there are no sessions. Claude Code:
claude mcp add --transport http topxai-models https://api.topxai.xyz/mcp/models \
--header "Authorization: Bearer $TOPXAI_API_KEY"
Cursor (~/.cursor/mcp.json) and other clients that take a JSON entry:
{
"mcpServers": {
"topxai-models": {
"url": "https://api.topxai.xyz/mcp/models",
"headers": { "Authorization": "Bearer sk-..." }
}
}
}
VS Code uses the same entry under servers with "type": "http". Add the images and videos servers the same way under their own names.
Limits
- The servers speak Streamable HTTP with one request per call:
POSTonly, no event stream, no sessions. A client sees changed prices or models the next time it lists the tools; a charge always uses the price at the time of the call. - There is no OAuth sign-in yet, so connectors that require OAuth (Claude.ai, ChatGPT) cannot connect. Developer clients that send a header can: Claude Code, Cursor, VS Code, Cline and curl.
- Without a key, or with a wrong one, a client can still connect and list the tools. A tool call then answers with what is wrong with the key, and nothing is sent to a model or charged.
- The
modelchoices ofask_modelfollow your key: a key on one route lists that route's models. - Image results up to 6 MiB are returned inline. Larger images come back as signed links valid for up to an hour; opening a link never starts another generation or changes the charge.
Related
- Hosted MCP servers
- Claude Code with TopxAI: ANTHROPIC_BASE_URL and the key
- How requests are billed
- Image generation
- Video generation
Previous: llm CLI with TopxAI: extra-openai-models.yaml · Next: Routes and the auto route
Other languages: 简体中文