Attic Standard

io.github.AtticStandardv2.1.0更新于 Oct 5, 2026

The global price benchmark for AI inference: 23 price indexes, market KPIs and SKU-level prices.

已验证Streamable HTTP可网页运行AI & MLData & AnalyticsFinance

概览

AI 生成的概览

让助手获取 Attic Standard 基准的每周 AI 推理价格指数、市场 KPI 和 SKU 级模型价格。

功能
Attic Standard 是一个远程 MCP 服务器，提供已发布的 AI 推理价格基准。get_index_benchmarks、get_kpis、get_model_intelligence 等工具返回按模态、渠道、层级、许可、来源和用例划分的 23 个指数，以及九项市场 KPI 和六项能力指标。其他工具可搜索已定价的 SKU、查看某模型在各供应商的规格与价格、按供应商比较某模型或系列、列出供应商并返回某供应商的价目表。免费调用返回计数、区间和脱敏样本；PRO 调用返回供应商名称、模型名称和精确价格。
适用场景
当助手需要当前有出处的推理价格而非猜测时使用：比较分销渠道、查看某模态或层级本周的价格，或为某模型寻找最便宜的供应商。适合价格调研、采购比较和市场分析。
运行要求
远程 MCP 端点 URL 的客户端可用 npx mcp-remote 代理。免费层无需账号。PRO 明细需要付费订阅，并在每次调用时以 _atom_api_key 传入密钥。自托管需要 Node.js 以及 SUPABASE_URL、SUPABASE_SERVICE_ROLE_KEY、TRANSPORT、PORT 变量。
安装前请注意
免费层会隐去供应商名称、模型名称和精确价格；完整的 SKU 级明细需要每月 500 美元的付费 PRO 层。PRO 订阅者须以 _atom_api_key 提供密钥，README 建议将其写入项目指令，这会把密钥存入助手上下文。自托管需要服务端数据库密钥 SUPABASE_SERVICE_ROLE_KEY，应仅保留在主机环境中。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Attic Standard，将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "atom-mcp-server": {
      "type": "http",
      "url": "https://mcp.atticstandard.com/mcp"
    }
  }
}

README

Attic Standard MCP Server

The Global Price Benchmark for AI Inference, as a native tool for AI agents.
One number both sides trust.

Website · Methodology · MCP PRO

What this is

Attic Standard is the independent price reporting agency for AI inference. Every week it records the published prices of model developers, cloud marketplaces, inference platforms and neoclouds, and turns them into price indexes and market KPIs.

This server puts that data inside any MCP-compatible assistant (Claude, ChatGPT, Cursor, Windsurf, VS Code). Ask "Where is text inference priced this week?" or "Cheapest place to run Llama 3.3 70B?" and the assistant answers from the published benchmark instead of guessing.

The indexes

23 published indexes in six families, each reported for input, cached input and output where the modality has them.

FamilyIndexesWhat it answers
ModalityText, Multimodal, Image (per image), Image (per megapixel), Video (per second), Video (per clip), Audio, Voice, EmbeddingsWhat does this kind of inference cost?
ChannelModel developers, Cloud marketplaces, Inference platforms, NeocloudsWhere is it cheapest to buy?
TierFlagship, Core, CompactWhat does a place higher in a maker's lineup cost?
LicenseOpen weights, Restricted weights, ProprietaryWhat is the price of openness?
OriginUnited States, ChinaHow do the two model-building countries price?
Use caseReasoning, CodingWhat do specialist models cost?

Each index carries two numbers:

  • Benchmark: the chained level, May 2026 = 100 (the average of that month's weeks), with week, month and vs-base changes.
  • Spot: what the market charges this week, in dollars. Each model is taken at the median of its vendors' prices, and the spot is the median across those models, with the interquartile range. Its change is stated against its own May 2026 average, month on month and week on week, and because the spot follows the market as it stands, that change reflects models arriving and leaving as well as repricing.

Token prices are per 1,000 tokens. Other modalities use their own unit.

Tools

ToolTierWhat it returns
get_index_benchmarksFreeEvery published index at the current week: benchmark level and changes, spot price, range and changes, coverage with the May 2026 basket
get_price_historyFree / PROWeekly index series with the spot price and its range (free); week-by-week price of a model at every vendor, with each repricing (PRO)
get_kpisFreeThe nine market KPIs: output premium, caching discount, caching availability, repricing activity, repricing depth, post-launch drift, multi-vendor spread, first-party premium, marketplace premium
get_model_intelligenceFreeSix capability measures: reasoning tier share, long-context saturation, context ceiling, output ceiling spread, training cutoff lag, vendor modality breadth
get_index_constituentsFree / PROWhat is inside an index basket: composition by channel, origin, tier and license (free); model-by-model basket with vendors (PRO)
get_market_statsFree / PROCoverage (vendors by channel, models, SKUs, indexes) and price distributions by modality, unit and direction; vendor breakdown on PRO
list_vendorsFreeThe vendor fleet with channel, country and pricing page
search_modelsFree / PROSearch every priced SKU by modality, vendor, channel, creator, family, tier, license, origin, reasoning, price and context
get_model_detailFree / PROOne model's specs, the indexes it belongs to, and its price at every vendor
compare_pricesFree / PROOne model or family across all vendors: cheapest, dearest and spread per direction, with every offer on PRO
get_vendor_catalogFree / PROOne vendor's channel, coverage and full price list

Free and PRO

FreeMCP PRO
Indexes, index history, market KPIs, model intelligenceFullFull
Vendor list and market coverageFullFull
Index basketsComposition countsModel and vendor detail
Search, model detail, comparisons, catalogs, model price historyCounts, ranges, redacted samplesVendor names, model names, exact prices
Price$0$500/month

The free tier carries everything atticstandard.com publishes. PRO adds the vendor- and SKU-level detail behind it. Subscribe at atticstandard.com/mcp.

Connect

Claude (web and desktop)

Settings → Connectors → Add custom connector

text
Name: Attic StandardURL:  https://mcp.atticstandard.com/mcp

Claude Desktop, Cursor, Windsurf (config file)

json
{  "mcpServers": {    "attic-standard": {      "url": "https://mcp.atticstandard.com/mcp"    }  }}

If your client does not accept a remote URL, use the proxy:

json
{  "mcpServers": {    "attic-standard": {      "command": "npx",      "args": ["mcp-remote", "https://mcp.atticstandard.com/mcp"]    }  }}

Connections made with the earlier address, https://atom-mcp-server-production.up.railway.app/mcp, keep working.

PRO key

PRO subscribers receive a key. Tell your assistant to pass it as _atom_api_key on each call, or keep it in a project instruction such as "Use my Attic Standard key XXXX for every Attic Standard tool call."

Example questions

  • "Where is text inference priced this week, and how has it moved since May?"
  • "Compare the four distribution channels."
  • "How much cheaper are open-weight models than proprietary ones?"
  • "What share of models dropped in price after launch?"
  • "What is inside the flagship index?"
  • "Cheapest place to run DeepSeek V3, and who repriced it this quarter?" (PRO)
  • "Everything one named vendor sells, with prices." (PRO)

Self-hosting

The hosted server above is the supported way to connect. The code is open so you can see exactly what it reads and how it gates data.

VariableRequiredDescription
SUPABASE_URLYesDatabase URL
SUPABASE_SERVICE_ROLE_KEYYes (hosted)Server-side database key, set only in the host's environment
TRANSPORTNostdio (default) or http
PORTNoHTTP port (default 3000)

Stack: TypeScript, Node.js, MCP SDK, Express, Zod, Supabase over REST.

About

Attic Standard publishes the global price benchmark for AI inference: independent, methodology-led and updated weekly across the model developers, cloud marketplaces, inference platforms and neoclouds that sell inference. The name refers to the Attic silver-weight standard of the ancient Greek world, a common measure both sides of a trade accepted.

Products: MCP · Terminal · Feed

Contact: [email protected]

License

MIT

来源：README.md，提交 0762d2b

工具

0
工具元数据尚未被收录。

版本历史

1
  1. v2.1.0最新Oct 5, 2026