If you use Claude Code, Cursor, Cline or VS Code with Copilot, you already have an MCP client. That means you can give your coding agent a live view of what AI models cost — including 111 Chinese models — without pasting a pricing page into the chat.
This guide covers the small set of MCP servers we publish for that, how to add them to each client, and what the returned data looks like. It is read-only and needs no API key, so you can wire it up in about two minutes.
Agent sessions usually start with a model choice: a cheap model for refactors, a stronger one for architecture, an image model for a mock-up. The decision is a lookup, not a conversation — exactly what a tool call is for. Instead of asking the model to remember prices (it will guess, and the guess will be stale), the agent asks the price feed at the moment of the decision.
The tool also normalises units: some models are priced per 1M tokens, some per generation, some per image or per second of video. Comparing raw numbers without units is how people accidentally pick an expensive model.
https://aiapi-pro.com/api/mcp. Three read-only tools: cheapest_models, model_price and list_categories. No auth required.io.github.vvvvking/novai-prices, and in punkpeye/awesome-remote-mcp-servers under Aggregators.The price data is also on the open web: /china-ai-leaderboard.json is regenerated daily from the live model catalogue, and /transparency.json states the fee and multiplier policy.
claude mcp add --transport http novai-prices https://aiapi-pro.com/api/mcpThen ask: “Which Chinese model is cheapest for a coding agent? Use the novai-prices server.”
// ~/.cursor/mcp.json
{
"mcpServers": {
"novai-prices": { "url": "https://aiapi-pro.com/api/mcp" }
}
}// .vscode/mcp.json
{
"servers": {
"novai-prices": { "type": "http", "url": "https://aiapi-pro.com/api/mcp" }
}
}MCP Servers → Remote Servers → add URL https://aiapi-pro.com/api/mcp. Leave headers empty.
{
"mcpServers": {
"novai-prices": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://aiapi-pro.com/api/mcp"]
}
}
}cheapest_models takes an optional category, a row limit and a price ceiling, and returns the cheapest models in that category with a comparable cost and the unit it was derived from. Illustrative shape:
{
"category": "Chat / LLM",
"models": [
{ "name": "GLM-5.3-Flash", "price_usd": 0.08, "unit": "flat", "url": "https://aiapi-pro.com/models/glm-5.3-flash.html" },
{ "name": "DeepSeek-V4-Flash", "price_usd": 0.21, "unit": "flat", "url": "https://aiapi-pro.com/models/deepseek-v4-flash.html" }
]
}model_price looks up a single model by name fragment; list_categories returns the available categories (Chat/LLM, Image, Video, 3D, Audio, Embedding). All three are read-only and return public data.
These are list prices as published by the upstream providers, passed through at a 1.0x multiplier with no platform fee. Units differ by model type — always read the unit column in the feed rather than comparing raw numbers.
| Category | Cheapest paid option | List price | Unit |
|---|---|---|---|
| Chat / LLM | Hunyuan MT2 Lite | $0.043 | flat |
| Image | Wan 2.2 T2I Flash | $0.0215 | per generation |
| Video | Seedance 2.0 Fast | $0.054 | flat |
| 3D | Hunyuan 3D Motion | $0.185 | per generation |
| Audio | Qwen3 TTS Flash | $0.0145 | flat |
| Embedding | Kinfra Text Embedding 0.6B | $0.0769 | flat |
The price server answers “what does it cost”. To actually run inference, use the OpenAI-compatible gateway at https://aiapi-pro.com/v1 with a key from /free:
from openai import OpenAI
client = OpenAI(base_url="https://aiapi-pro.com/v1", api_key="nvai-...")
resp = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[{"role": "user", "content": "Explain MCP in one sentence."}],
)
print(resp.choices[0].message.content)Because the wire format is OpenAI Chat Completions, existing SDK code and agent frameworks work with a base-URL change. The same key serves text, vision, image and video models; billing follows the unit shown in the feed.
That is the whole point of exposing prices as a tool: the agent can make the cost decision per step, with current numbers, instead of a hard-coded guess.
No. The price tools are read-only and public: point any MCP client at https://aiapi-pro.com/api/mcp. You need a key only when you call the models for inference.
Any client that supports remote Streamable HTTP MCP servers: Claude Code, Cursor, VS Code with Copilot, Cline, and Claude Desktop via mcp-remote. Config snippets for each are in the article.
Yes. NovAI bills paid models at the upstream provider list price with a 1.0x multiplier and a 0% platform fee. The policy is published in /transparency.json and the per-model feed is /china-ai-leaderboard.json.
Five GLM Flash models (text, vision, image, video), 10 free-model calls per day and 5 requests per minute, with no credit card required.
Daily, from the live model catalogue. Each entry carries its own unit (per 1M tokens, per generation, per image, flat), because units differ by model type.
One OpenAI-compatible endpoint for DeepSeek, Qwen, GLM, Kimi, MiniMax and Doubao. Five models are free to call without a credit card.
Get a free API key →