NanoParse

NanoParse MCP

[npm version] [License: MIT] [Website] [MCP Score] [MCPLookup Trust Index] [Listed on HookMCP]

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" },    "freshness": { "published": null, "age": "unknown", "stale_warning": "No publish date found — verify independently", "modified": null, "has_correction": false, "correction_note": null },    "tldr": "Abstract page for arXiv paper 2401.14295: Demystifying Chains, Trees, and Graphs of Thoughts",    "key_figures": [ { "value": "2,247", "context": "Thu, 25 Jan 2024 16:34:00 UTC (2,247 KB)" } ],    "structural_trust_score": { "level": "medium", "trust_factors": ["High-authority source", "No publish date found — verify independently"] },    "hedge_language": { "density": "low", "markers_found": 0, "per_500_words": 0 },    "content_density": { "thin_content": false, "substantive_sentence_ratio": 0.76 },    "syndication": { "detected": false, "note": null },    "paywall": { "detected": false, "note": null },    "content_type": "academic_paper",    "reading_time": { "minutes": 4 },    "is_archived": { "archive_org": true, "archive_is": false },    "outbound_authority": { "score": 0.54, "top_domains": ["ui.adsabs.harvard.edu", "scholar.google.com"] },    "boilerplate_ratio": { "ratio": 0.08, "confidence": "high" }  }}

The agent learns before reading a word: this is an academic source (0.85 authority), structural trust is medium — high authority, but no publish date found, so verify independently. It also gets the one-sentence summary, the key numbers, and confirmation there's no paywall or syndication flag. See examples/agent-conversation.md for a full worked round-trip.

Payments

• 10 free parses per month with a signed wallet (sign the claim message — no payment), or 3 anonymous free tries per network. No account, no API key, no human required.
• After that: $0.005 per parse (flat, no subscriptions) via x402 — an open micropayment protocol. Your agent's wallet signs the payment automatically. To give your agent a Coinbase wallet, run npx @coinbase/payments-mcp and connect it alongside NanoParse.

Copy-paste recipe for both paths — the wallet-MCP one with no code, and a ~40-line REST client — including the exact 402 → sign → retry loop and the error codes worth branching on: examples/pay-with-x402.md.

How payments work under the hood

When an agent without free quota calls nanoparse_fetch, the endpoint responds with an HTTP 402 — Payment Required carrying x402 payment instructions. The agent's wallet signs a $0.005 USDC (Base) transfer, the facilitator settles it, and the parse proceeds. The hosted endpoint handles the entire challenge → settlement flow — the agent never needs a private key on your machine.

Pricing vs Firecrawl

Full math, including when Firecrawl's base rate wins at scale: examples/costs.md.

Example

You: Fetch https://arxiv.org/abs/2401.14295 and summarize the argument.
Agent: [calls nanoparse_fetch(url)]       → clean Markdown + Litmus (academic, authority 0.85, 4-min read,         no date found — verify, no paywall)       → summarizes from primary content, cites authority and freshness

Repository note

The nanoparse-mcp npm package and the local stdio client were the first-generation integration, which held a wallet private key (NANOPARSE_WALLET_KEY) and signed x402 payments on your machine. Both are retired: the hosted endpoint above now handles the full payment flow itself, so no local client or wallet key is needed. The retired client code is preserved in archive/v1-local-client/ for history, and the npm package is deprecated. Use the hosted endpoint — it is the current, supported integration.

Docs

• Quick start & integration: nanoparse.app/quickstart
• Compare & pricing: nanoparse.app/compare
• API + MCP endpoint: nanoparse.app
• Security policy: SECURITY.md — report a vulnerability to [email protected] (also published at https://nanoparse.app/.well-known/security.txt)

MIT licensed.