AI代理读取网页的最简单方式。 NanoParse MCP是一个托管的模型上下文协议服务器，位于https://nanoparse.app/mcp，将任意URL转换为智能体实际可用的干净Markdown——并附加机器可读的信任信号。

无需安装。无需钱包密钥。无需.env。连接一次，您的代理就能访问网页。

为什么代理选择NanoParse而不是Firecrawl或Jina Reader： 每个页面都会带回Litmus——20个机器可读信号，告诉你的代理是否信任并引用该页面，然后才读到一个单词。免费套餐结束后，通过x402每次解析支付0.005美元。完整的JavaScript/SPA渲染，原生GFM表。无账户，无订阅，无API密钥。

快速开始

把这个添加到你的MCP客户端配置里：

重新启动客户后，你的经纪人会自动发现两个工具：

• nanoparse_fetch(url) — 获取任意网页并恢复干净的Markdown
• nanoparse_status() — 消费前请核查免费配额、钱包和USDC余额

仅此而已。你的代理现在可以从任何兼容MCP的客户端（Claude、Cursor、Hermes、Windsurf等）中获取页面。

你得到了什么

NanoParse MCP

The easiest way for AI agents to read the web. NanoParse MCP is a hosted Model Context Protocol server at https://nanoparse.app/mcp that turns any URL into clean Markdown your agent can actually use — with machine-readable trust signals attached.

No install. No wallet key. No .env. Connect once and your agent can read the web.

Why agents choose NanoParse over Firecrawl or Jina Reader: every page comes back with Litmus — 20 machine-readable signals that tell your agent whether to trust and cite the page before it reads a word. Pay $0.005 per parse via x402 once the free tier runs out. Full JavaScript/SPA rendering, native GFM tables. No account, no subscription, no API key.

Quick start

Add this to your MCP client config:

{
  "mcpServers": {
    "nanoparse": {
      "type": "http",
      "url": "https://nanoparse.app/mcp"
    }
  }
}

Restart your client, and your agent auto-discovers two tools:

• nanoparse_fetch(url) — fetch any web page and get clean Markdown back
• nanoparse_status() — check free quota, wallet and USDC balance before spending

That’s it. Your agent can now fetch pages from inside any MCP-compatible client (Claude, Cursor, Hermes, Windsurf, and others).

What you get

nanoparse_fetch renders the page in a real headless browser (JavaScript and all), isolates the actual article from nav, ads, and cookie banners, and returns:

• Clean Markdown — GFM tables, callouts, math, and footnotes preserved
• Metadata — title, author, dates, language, word count, OpenGraph, schema.org JSON-LD
• Litmus — machine-readable signals your agent can reason over: source authority, freshness, structural trust, content density, syndication and paywall detection

nanoparse_status tells your agent exactly where it stands: free parses remaining, wallet and USDC balance, and whether the next call will succeed without payment. Agents call it before spending.

Real output

Every parse returns markdown + metadata + litmus. Here is the full litmus object from a real parse of an arXiv paper (examples/example-output.json — markdown truncated there, litmus complete and unmodified):

{
  "litmus": {
    "source": { "type": "academic", "authority": 0.85, "rationale": "academic source" },