@modelpricewatch/mcp

MCP server for live LLM API pricing — current input/output token prices, side-by-side comparisons, and cheapest-model lookups for 150+ models across 20+ providers (OpenAI, Anthropic, Google, Mistral, xAI, DeepSeek, …), re-verified against official provider pricing pages by modelpricewatch.com.

Unlike MCP servers that bake a pricing snapshot into each release, this one is a thin bridge to a live API — prices are never stale between package releases, and every price links back to the provider's official pricing page.

Tools

Tool	What it answers
search_models	"How much does model cost?" / "What models does provider offer?"
get_model_pricing	Full pricing + capabilities for one model (input/output/cached $/1M tokens, context, modality)
compare_models	2–5 models side by side, with a cheapest-verdict
cheapest_models	Cheapest models ranked by input/output/blended cost, filterable by category / open-source
list_providers	All tracked providers with model counts

Setup

Claude Code (HTTP transport — no install needed):

claude mcp add --transport http modelpricewatch https://modelpricewatch.com/mcp

Claude Desktop / stdio-only clients (claude_desktop_config.json):

{
  "mcpServers": {
    "modelpricewatch": {
      "command": "npx",
      "args": ["-y", "@modelpricewatch/mcp"]
    }
  }
}

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "modelpricewatch": { "url": "https://modelpricewatch.com/mcp" }
  }
}

How it works

This package is a zero-dependency stdio↔HTTPS bridge to https://modelpricewatch.com/mcp (streamable-http, stateless). No API key required. Data is CC-BY: cite modelpricewatch.com. The underlying JSON API is also free and CORS-open — see modelpricewatch.com/api.

Fair-use allowances

Expressed per minute, not per day, because agent traffic is bursty: a daily quota either cuts a live conversation short or does nothing about a client stuck in a retry loop.

These are allowances to design against, not hard caps. Tool-call metering is deliberately loose and will usually let a burst through; sustained excess may be blocked at the network edge without warning.

Plan	Tool calls / minute	Counted per
Free (no key)	20	IP address
Pro	240	API key
Startup	2,400	API key
Enterprise	unmetered	API key

Connecting, listing tools and pinging are counted separately at 120/min per anonymous IP, so reconnecting never spends your tool-call budget.

Exceeding an allowance returns a normal MCP tool result flagged isError — not a transport error — carrying Retry-After and an explanation of how to raise the ceiling, so the assistant can tell you what happened.

Do not poll the discovery endpoint

GET https://modelpricewatch.com/mcp returns the discovery document. It changes only when the site is deployed, so fetch it once per session and cache it — sustained polling may be throttled at the edge. It returns an ETag and Cache-Control: max-age=300; a conditional request with If-None-Match gets a 304. If you need a liveness check, every few minutes is ample — a short timer buys you nothing extra.

Normal tool use is POST and is unaffected by that limit.

To raise the ceiling, send your key as X-API-Key or Authorization: Bearer (set it in the env block of your MCP client config). Keys are issued by request: modelpricewatch.com/contact. Plans: modelpricewatch.com/api-pricing.

The pricing data itself is free and stays free — a paid tier buys throughput, a commercial license and an SLA, not access to the numbers. If you are being rate limited on bulk reads, the REST API serves the same dataset with no key at all.

License

MIT (the bridge). Pricing data: free with attribution — modelpricewatch.com.