Apipriceindex MCP

apipriceindex-mcp

An MCP server that gives any AI agent live, verified LLM API prices — and today's deals.

Ask your assistant *"which LLMs are free to call right now?"*, *"what's the cheapest model at least as good as GPT-5.6 Sol?"* or *"what would 200M input / 15M output tokens a month cost me on Sonnet vs GPT?"* and it answers from today's index — every price with its verification date, cross-check confidence and source URL. Data from apipriceindex.com (CC BY 4.0, re-checked daily from each provider's own price list or from OpenRouter's per-provider data; every price names its source in source_url).

Tools

Install

pip install apipriceindex-mcp

Claude Code / Claude Desktop:

claude mcp add apipriceindex -- apipriceindex-mcp

Or in any MCP client config:

{ "mcpServers": { "apipriceindex": { "command": "apipriceindex-mcp" } } }

Behavior

• Data is fetched live from the index (15-minute in-memory cache) — never stale training data.
• Index unreachable → explicit error. No cached or invented price is ever served silently.
• Index older than 14 days → answers still come, with a warning pointing at the index's health endpoint.
• Model not tracked → says so, links the dataset. No price invented.
• Cache discount only where the provider publishes a cached-input price.
• Deals always carry an end date — no end date when the host publishes none — and the link to the host's own page.
• Speed and availability are read from OpenRouter and always come with their window (median over 30 minutes, share of successful requests over 24 hours). No tail latency or error rates.
• What a host serves (weights precision, context, max output) is read at the host's own catalogue or docs where its terms allow, otherwise as the host declares to OpenRouter. Each value carries its state: published, not_stated (the host's source does not state it), not_read (the host's terms forbid automatic reading and OpenRouter has no value) or to_check. Nothing is guessed.
• One third-party score: the Epoch Capabilities Index by Epoch AI (CC BY 4.0). Every answer that carries it carries its credit. Otherwise, cost only: the index does not rank model quality.

Environment: APIPRICEINDEX_URL overrides the dataset URL (testing); deals.json and pricing-history.csv are read next to it.

Changelog

• 0.3.0 — What each host serves of an open-weight model: get_price returns served.precision, served.context and served.max_output, each with its state, source and date, and the maker's values in served.maker; offers_table takes precision and max_output columns, and an empty served cell comes with _state. Other answers unchanged.
• 0.2.1 — get_price and this README now say where prices are read: each provider's own price list, or OpenRouter's per-provider data, named by source_url. They said "official pricing pages", which was true of about one price in four. No change to the tools or their answers.
• 0.2.0 — list_deals, offers_table, compare_models, cheapest_at_level, price_changes; search_models filters by max price, capabilities and min speed; Epoch score, speed and availability where known. The three 0.1 tools answer as before.
• 0.1.1 — search_models, get_price, estimate_cost.

License

Code: MIT. Price data: CC BY 4.0, attribution "API Price Index". Epoch Capabilities Index: CC BY 4.0, credit Epoch AI.