io.github.vvvvking/novai-prices-mcp
<!-- mcp-name: io.github.vvvvking/novai-prices-mcp -->
# novai-prices-mcp
**Listed in the [official MCP Registry](https://registry.modelcontextprotocol.io) as `io.github.vvvvking/novai-prices`** — remote `streamable-http` endpoint: `https://aiapi-pro.com/api/mcp`?utm_source=github_readme&utm_medium=readme
An MCP server that answers one question well: **which Chinese AI model is cheapest for what I'm doing right now?**
It reads NovAI's live price feed (`https://aiapi-pro.com/china-ai-leaderboard.json`?utm_source=github_readme&utm_medium=readme) and normalizes every model into a comparable cost unit, because every provider quotes a different one:
| Modality | Unit the market quotes | Unit this server uses |
|---|---|---|
| Chat / LLM | per 1M input, per 1M output | **1M input + 200K output** (one large coding run) |
| Video | per second, per generation, per 1M tokens | **one 5-second clip** (token-billed models are converted) |
| Image | per image, per generation | **one image** |
Prices are the providers' list price passed through with a **1.0 multiplier and 0% platform markup** — the server reports the rate card, not a sales number.
## Tools
| Tool | What it does |
|---|---|
| `cheapest_models` | Cheapest models, optionally filtered by `category` and `max_price` |
| `model_price` | Price lookup by name fragment (`seedance`, `kimi`, `glm`, `qwen`…) |
| `list_categories` | Categories available in the feed |
## Run it
No dependencies beyond Python 3.9+:
```bash
python3 server.py
```
### Claude Desktop / Cline / any MCP client
```json
{
"mcpServers": {
"novai-prices": {
"command": "python3",
"args": ["/absolute/path/to/server.py"]
}
}
}
```
### Example session
```
→ tools/call cheapest_models {"category": "video", "limit": 3}
← CogVideoX-Flash (Zhipu AI) — FREE
PixVerse C1 Video API (PixVerse) — $0.154 per 5s clip (est.)
PixVerse V6.0 Video API (PixVerse) — $0.203 per 5s clip (est.)
```
```
→ tools/call cheapest_models {"category": "chat / llm", "limit": 3}
← GLM-4.7-Flash — FREE
DeepSeek V4 Flash — $0.080/1M in, $0.170/1M out ($0.11 per 1M-in/200K-out run)
GLM-5.3-Flash — $0.080/1M in, $0.280/1M out
```
## Why this exists
Price comparison for Chinese models is broken: one vendor quotes per second, another per token, a third per generation. Converting them by hand is error-prone, and every gateway's marketing page has an incentive to make its own numbers look good. This server just converts and sorts, and links to the machine-readable feed so the numbers can be checked.
## Data freshness
The feed is regenerated daily from the live model catalog. `generated_at` is included in the JSON; if the upstream feed is unreachable the tool returns an error rather than stale numbers.
## License
MIT
## Where it is listed
- Official MCP Registry: `io.github.vvvvking/novai-prices` (streamable HTTP)
- Glama connector: https://glama.ai/mcp/connectors/io.github.vvvvking/novai-prices — TDQS 3.6/5, Healthy
- `punkpeye/awesome-remote-mcp-servers` (Aggregators)
- mcpservers.org: https://mcpservers.org/servers/novai-prices-mcp (12 language versions)
- mcpub.dev: https://mcpub.dev/?q=novai
---
## Data source
The price data comes from NovAI's public price index: https://aiapi-pro.com/chinese-ai-api-pricing - one page per billing unit, generated from the same `/v1/models` endpoint the API serves. Every number this server returns can be checked against that page.
- Coverage: 114 models across 22 Chinese providers (measured from `GET https://aiapi-pro.com/v1/models`).
- Unit normalisation: prices are per million tokens, except media models (video / image / 3D / speech / music), which are billed per generation. The server keeps those in separate categories instead of mixing units.
## Update frequency
The index is regenerated weekly. `data/*.json` in this repository is the snapshot the server reads; if a price changes mid-week the snapshot is updated and **the commit timestamp on that file is the authoritative "as of" date**.
## Maintenance
Maintained by the NovAI team (https://aiapi-pro.com). Issues and pull requests are welcome for: wrong prices (include the vendor's own source URL), missing models, and normalisation errors. Expect an answer within a few days - this is a side project of the same team that runs the API, not a support channel.
## Not a promotion
The purpose of this server is to answer "which Chinese model is cheapest for this workload" with numbers that trace to a source and a date. It is not an advertisement: if the data is wrong, the correct outcome is a corrected dataset, not a better claim. Discounts and free tiers are reported because they change the answer, not because they are flattering.
TDQS
Scored across 3 tools
The three tools are largely distinct: model_price is a targeted name-fragment lookup, cheapest_models is a ranked browse of the whole feed, and list_categories returns taxonomy only. The only mild overlap is that model_price and cheapest_models both surface price data, but the descriptions make the lookup-vs-ranking distinction clear.
All names are snake_case, which is good, but the grammatical pattern is mixed: model_price is a bare noun, cheapest_models is adjective_noun, and list_categories is verb_noun. It is readable but does not follow a single predictable convention.
Three tools is lean but reasonable for a read-only price-feed server covering lookup, ranking, and category enumeration. Nothing is redundant, though the surface is close to the thin end for anything beyond simple queries.
Core pricing queries are covered, but there is a notable gap: list_categories returns categories that no other tool can use as a filter, so an agent cannot list or rank models within a category. There is also no way to enumerate the full catalog or compare several specific models side by side.