Your agent deservesbetter data.
Every parse returns clean Markdown plus Litmus — twenty deterministic signals on who wrote the page, how fresh it is, and whether it is even the page you asked for.
10 free parses a month · no account · no API key
Your agent can read any page.It cannot tell whether to believe it.
Extraction solved the text problem and left the judgement problem alone. Each of these is something an agent gets wrong today — confidently, and in writing.
Stale, cited as current
No publish date, no staleness flag. A 2019 explainer reads exactly like last week's — and the citation looks identical either way.
A paywall read as content
A login wall or truncated teaser passes as the article. The agent summarises the first eighty words of a page it never actually read.
An error page read as content
A soft 404, a captcha, a "please enable JavaScript" shell — all arrive with a 200 status and get treated as a source.
A copy read as the original
Thin pages, syndicated reposts and boilerplate-heavy layouts rank like primary sources, because nothing measured the difference.
Twenty data points before your agent cites a page.
Every parse returns clean Markdown plus Litmus — twenty rule-based checks on who wrote the page, how fresh it is, and whether it is even the page that was asked for. No model in the path.
Source
6 signalsWho wrote it, what it claims to be, who it links to, and whether it shows its work.
Page type, authority score, and the reasoning behind it.
type: "news" · authority: 0.82type: "blog" · authority: 0.21No model in the path.Same page, same verdict.
A model can form an opinion about a page. The problem is what happens on the second run. Litmus is rule-based and extractive, so a given page and version produce the same signals every time — and the verdict can be logged and audited.
“A judge that changes its mind is not a judge you can log.”
A model in the path
- Different verdict on the same page tomorrow
- Unbounded cost per call, paid in tokens
- No audit trail — "the model said so"
- Confident when wrong, with no field to check
Determinism in the path
- Same signals on the same page, always
- Fixed cost — $0.002, metered per call
- Every field inspectable in the payload
- Fails visibly — null and "unknown", never a guess
x402. USDC on Base.$0.002 per parse.
No account, no API key, no subscription. An agent discovers the endpoint, pays per parse, machine to machine, and is charged only when the page renders.
On a cloud host the anonymous allowance does not apply: sign with your wallet once and claim the same 10, no payment and no account. How the claim works
One config. Then your agent has the web.
No SDK, no build step. Add the MCP server and the agent discovers both tools — fetch a URL, or check its quota and wallet state.
- MCP
nanoparse_fetch(url) → markdown + litmus - MCP
nanoparse_status() → quota, wallet, payment - REST
POST /fetch · OpenAPI at /openapi.json
{ "mcpServers": { "nanoparse": { "type": "http", "url": "https://nanoparse.app/mcp" } }}// The agent discovers two tools:// nanoparse_fetch(url) -> markdown + litmus// nanoparse_status() -> quota, wallet, paymenttypically 2–8 seconds per parse · results held 6 hours
Other tools extract. Litmus judges.
Questions agents would ask.
Straight answers for developers wiring up NanoParse.
Yes. Standard scrapers fail on modern SPAs because they only fetch raw HTML. NanoParse renders each page in headless Chromium, executes JavaScript, and waits for dynamic content to settle before running content isolation. The Markdown you get back is what a human would have seen.
Test the page. Then trust it.
Run it on a page you already trust, then on one you don't. The difference in the payload is the pitch.
10 free parses a month · no account · no API key