web content extraction for autonomous agents

Any URL in. Clean, structured content out.

Name: CleanPage
Price: 0.02 USD

CleanPage is an MCP tool that strips a page down to what an agent actually needs — readable markdown, metadata, links, and structured facts — with a quality score attached, so your agent never has to guess whether the output is any good.

Get MCP connection details See pricing

$0.02 USDC per call x402 on Base no signup, no API key

clean_page — live preview

Output will appear here. This preview truncates to 700 characters — the real clean_page tool returns full markdown, metadata, links, images, and structured facts.

Free preview, rate-limited. The MCP tool below returns the untruncated result.

output shape

One call, everything an agent needs to reason about a page

No HTML to parse, no boilerplate to strip, no guessing whether the page loaded correctly. Every response is the same predictable JSON shape.

clean_markdown / plain_text

Article body only — nav, ads, cookie banners, and related-article rails are already gone. Markdown preserves headings, lists, links, and tables; plain_text is the flat version.

metadata

Author, published date, site name, language, word count, and estimated reading time, pulled from the page's own byline and meta tags where available.

structured_facts

Optional array of standalone claims, entities, statistics, quotes, and dates, each with a confidence score — useful for downstream retrieval or fact-checking.

links & images

Internal and external links with their anchor text, plus resolved absolute image URLs, deduplicated and capped to a sane count.

quality_score

A 0–100 heuristic score based on content length, structure, and metadata completeness — so an agent can decide whether to trust or discard a result programmatically.

extraction_notes

Plain-language notes about anything unusual (short content, truncation, skipped enrichment) instead of failing silently or hallucinating a clean result.

pricing

Pay per call. Nothing else.

No subscription, no minimum spend, no dashboard to sign into. Failed extractions — paywalls, bot-blocks, empty pages — are never charged.

$0.02USDC / call

clean_page — the only paid tool. Everything else on this server is free.

Charged only on successful extraction
Settled on Base via x402, USDC
No account, API key, or wallet whitelisting
health, list_capabilities, pricing are free, always

Read the API reference

Call `clean_page` with no payment

You get back HTTP 402 with an accepts array describing exactly what payment is required — amount, network, asset, and payee address.

Sign a USDC authorization

Construct an EIP-3009 transferWithAuthorization payload for the exact amount requested — most x402 client libraries do this automatically from a wallet key.

Retry with `X-PAYMENT`

Same request, same tool call, with the base64-encoded payment payload attached as a header. The server verifies it against the facilitator before doing any work.

Get your result, plus a receipt

On success, the response carries an X-PAYMENT-RESPONSE header with the settlement transaction hash. If extraction fails, settlement never happens — you keep your USDC.

for agents

Connect in one config block

CleanPage speaks MCP over Streamable HTTP. Point any MCP-compatible agent framework at the endpoint below.

// mcp client config
{
  "mcpServers": {
    "cleanpage": {
      "url": "https://cleanpage-mcp.cleanpage-mcp.workers.dev/mcp",
      "transport": "streamable-http"
    }
  }
}

Discovery, without a human in the loop

Every capability and price is machine-readable before you spend anything. Call list_capabilities or pricing over MCP, or fetch the static files below directly.

/.well-known/agent.json /.well-known/mcp.json /llms.txt /pricing /health

Example tool call

A minimal tools/call request. The first attempt with no payment returns HTTP 402; retry with the resulting X-PAYMENT header to receive the actual result.

POST /mcp
{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "clean_page",
    "arguments": {
      "url": "https://example.com/article",
      "options": { "extract_facts": true }
    }
  }
}

error handling

Typed errors instead of silent junk

If a page can't be cleanly extracted, you get a clear reason and whether it's worth retrying — never a charge for a failed call.

error_code	meaning	retryable
INVALID_URL	URL is malformed, non-http(s), or resolves to a blocked internal address.	no
FETCH_FAILED	The site returned a non-2xx status, non-HTML content, or was too large.	sometimes
TIMEOUT	The page didn't respond within the configured timeout.	yes
PAYWALL_DETECTED	The page is behind a subscription/metered paywall; full content isn't accessible.	no
BLOCKED_BY_SITE	Bot-check, CAPTCHA, or a JS-only shell with no server-rendered content.	no
EMPTY_CONTENT	No substantial article content was found (listing page, homepage, etc).	no
EXTRACTION_FAILED	The extraction engine hit an unexpected parsing error.	no