AI Price Index MCP server

ai-price-index-mcp

RoninForge ai-price-index-mcp is a read-only Model Context Protocol server that answers what an AI model's API price was on a given date, with the first-party source that proves it. It is MIT licensed, runs entirely on your machine, and requires no API keys.

Part of RoninForge.org, an independent open-source workshop that keeps dated, reproducible records of the AI developer tooling economy. It serves the open AI Price Index dataset to any MCP client, including Claude Code and Cursor.

Most pricing tools tell an agent what a model costs today. This one also answers what a model cost on a given date, and returns the first-party source URL plus the date that price was last validated, so the agent can cite the number. That point-in-time lookup is the differentiator.

ai-price-index dependency. The server reads it locally and adds nothing to your traffic.

• Zero keys, zero prompts, zero network at runtime. The dated dataset is bundled inline inside the

attribution string.

• Read-only. It only looks prices up. It cannot write, configure, or call out.
• Cite-ready. Every result carries sourceurl, lastvalidated, confidence, and the CC BY 4.0

Install

Requires Node.js 18 or newer. No build step, no API key.

npx -y ai-price-index-mcp

That command runs the stdio server; an MCP client launches it for you using the config below.

Client configuration

Drop this into your MCP client config. The server speaks stdio.

Claude Code (~/.claude.json or a project .mcp.json), and Cursor (~/.cursor/mcp.json), use the same mcpServers shape:

{
  "mcpServers": {
    "ai-price-index": {
      "command": "npx",
      "args": ["-y", "ai-price-index-mcp"]
    }
  }
}

Or, with Claude Code's CLI:

claude mcp add ai-price-index -- npx -y ai-price-index-mcp

Tools

All tools return structured JSON as text content. Prices are usdpermtok (USD per million tokens). Model ids resolve through the dataset's aliases, so short ids work (e.g. claude-opus-4-5 resolves to claude-opus-4-5-20251101).

tokens for costfromusage accepts input, output, cacheread, cachewrite5m, cachewrite_1h (all optional, missing counts as 0). Cache read is 0.1x input, cache write is 1.25x (5 minute) or 2x (1 hour).

Pass provider (for example openai, anthropic, google) to disambiguate a bare id that exists under more than one vendor.

Example result

price_on with { "model": "gpt-4", "date": "2024-01-01" }:

Shortened. The full README is on GitHub.