PDF to Markdown API

Converts a PDF to Markdown, inferring headings from font size (best-effort, not layout-perfect; no OCR).

Updated

Agents read this page asGET /v1/pdf-to-markdown/schemafree: schema, examples, errors, price

Endpoint
POST /v1/pdf-to-markdown
Price
$0.005 per call
Timeout
40 s
max_timeout_seconds
61 (the whole call as the 402 envelope sells it: the timeout, plus payment verification and settlement around it)
Response cap
25 MB (26,214,400 bytes)
Family
Documents

When to use it

A PDF has no heading structure, only sized glyphs; recovering Markdown needs a real PDF engine plus a hardened fetch, which an agent cannot do inline without walking past the SSRF controls. Offline, best-effort text-layer extraction; no OCR. No external CMap or standard-font data is configured or fetched. PDFs with self-contained character mappings can yield text, including CJK; documents requiring unavailable predefined CMaps can return empty or partial text without an error. Embedded ToUnicode alone does not guarantee extraction when the encoding still needs an external CMap. Missing standard-font data does not necessarily prevent extraction. A successful response does not certify complete text; truncated reports only the max_pages limit.

Your source is PDF and you need Markdown.

When not to use it

  • The input is over 25 MB (26,214,400 bytes): pdf-to-markdown answers too_large (413).
  • The destination needs more than 40 s to answer: the call ends with upstream_timeout (424).
  • The destination is on a private or local network: the network guard refuses it with blocked_target (403).

Parameters

One JSON body. The schema is free at GET /v1/pdf-to-markdown/schema.

FieldTypeRequiredDescription
urlstringyesAbsolute http(s) URL of the PDF, up to 2048 characters; redirects are followed and the body must start with the %PDF- signature.
max_pagesintegernoHow many leading pages to convert, 1 to 1000; a shorter document is not an error, and a longer one is cut and reported as truncated. Default: 100.

Limits

LimitValue
Largest source file25 MB (26,214,400 bytes)
Timeout40 s
Source formatsPDF
Output formatsMarkdown
Show the JSON Schema the handler validates against
JSON Schema
{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "type": "object",
  "properties": {
    "url": {
      "type": "string",
      "maxLength": 2048,
      "format": "uri",
      "description": "Absolute http(s) URL of the PDF, up to 2048 characters; redirects are followed and the body must start with the %PDF- signature."
    },
    "max_pages": {
      "default": 100,
      "description": "How many leading pages to convert, 1 to 1000; a shorter document is not an error, and a longer one is cut and reported as truncated.",
      "type": "integer",
      "minimum": 1,
      "maximum": 1000
    }
  },
  "required": [
    "url"
  ],
  "additionalProperties": false,
  "description": "Offline, best-effort text-layer extraction; no OCR. No external CMap or standard-font data is configured or fetched. PDFs with self-contained character mappings can yield text, including CJK; documents requiring unavailable predefined CMaps can return empty or partial text without an error. Embedded ToUnicode alone does not guarantee extraction when the encoding still needs an external CMap. Missing standard-font data does not necessarily prevent extraction. A successful response does not certify complete text; truncated reports only the max_pages limit. PDF text extraction accepts at most 1000000 raw and separately 1000000 output UTF-16 code units, 10000 items per page and 100000 items across selected pages. Exceeding a limit rejects the call with 413 too_large. truncated refers only to max_pages."
}

Example

The published example, verbatim: the request and the response it returns.

Request
POST /v1/pdf-to-markdown
content-type: application/json
payment-signature: <base64 x402 payload>

{
  "url": "https://grist.tools/samples/documents/markdown-guide.pdf",
  "max_pages": 100
}

Without the payment-signature header the call answers 402 with the payment requirements, including this endpoint's price and its max_timeout_seconds, so the deadline is known before anything is paid. The payment is verified before the handler runs and settled only after it succeeds, so a call refused by the gate, or one that fails inside the handler, is never settled. The single case we cannot state for you is a settlement call that fails without answering: the error message says so when that is what happened.

Response200 OK
200 OK
payment-response: <base64 settlement receipt>

{
  "source_url": "https://grist.tools/samples/documents/markdown-guide.pdf",
  "total_pages": 1,
  "extracted_pages": 1,
  "truncated": false,
  "chars": 249,
  "markdown": "# Field Notes\n\n## A short guide in three parts\n\nThis sample shows how headings are inferred.\n\nLarger text becomes a heading, body text stays plain.\n\n### Second section\n\nEach line of the page becomes its own block.\n\nThe document ends after this line."
}

Sample file: documents/markdown-guide.pdf. One-page PDF with lines drawn at 24, 18, 14 and 12 pt Helvetica, so a 12 pt body sits under three heading sizes; built with pdf-lib with pinned metadata.

In the sample itself, the PDF 1.7 file has 3 top-level objects, 1 object stream among them, and a cross-reference stream; inflated, its streams declare 1 page with MediaBox 0 0 420 320, base font Helvetica and 7 text runs at 24, 18, 12 and 14 pt, the first reading "Field Notes". The returned Markdown splits into 7 blocks separated by blank lines, with 3 headings (1 at level 1, 1 at level 2, 1 at level 3) reading "Field Notes", "A short guide in three parts" and "Second section". Plain paragraphs make up 4 blocks, the first reading "This sample shows how headings are inferred." and the last "The document ends after this line." Its headings total 10 words, its paragraphs 31 words.

Response fields

FieldTypeIn the example
source_urlstring"https://grist.tools/samples/documents/markdown-guide.pdf"
total_pagesnumber1
extracted_pagesnumber1
truncatedbooleanfalse
charsnumber249
markdownstring"# Field Notes ## A short guide in three parts This sample shows how headings are inferred. Larger text becomes a heading, body text stays plain. ### Second section Each line of the page becomes its own block. The document ends after this line."

Deterministic: the same input returns the same output. Clock-derived fields, when present, are declared in the response.

Call it from code

The client reads the 402, signs the payment and retries on its own.

// pdf-to-markdown: $0.005 per call, paid in USD Coin on eip155:8453 via x402
// Install: npm install @x402/fetch@2 @x402/evm@2 viem@2
// Save as client.mjs (ES module, Node 18+), export EVM_PRIVATE_KEY with the paying wallet's key in your shell, then run: node client.mjs
// The wrapper reads the 402 (payment-required), signs and retries with payment-signature.
import { wrapFetchWithPayment, x402Client, decodePaymentResponseHeader } from '@x402/fetch'
import { ExactEvmScheme } from '@x402/evm/exact/client'
import { privateKeyToAccount } from 'viem/accounts'

const account = privateKeyToAccount(process.env.EVM_PRIVATE_KEY)
const client = new x402Client().register('eip155:8453', new ExactEvmScheme(account))
const fetchWithPayment = wrapFetchWithPayment(fetch, client)

const body = {
  "url": "https://grist.tools/samples/documents/markdown-guide.pdf",
  "max_pages": 100
}

const res = await fetchWithPayment('https://grist.tools/v1/pdf-to-markdown', {
  method: 'POST',
  headers: { 'content-type': 'application/json' },
  body: JSON.stringify(body),
})
console.log(res.status, await res.json())
const settle = res.headers.get('payment-response')
if (settle) console.log(decodePaymentResponseHeader(settle).transaction)

Errors

Every error is typed JSON. Only a successful call is settled.

CodeHTTPWhen
invalid_input400The body does not match the schema on this page. Nothing is charged. The body does not match the schema; its fields are url and max_pages.
blocked_target403The destination is refused by the network guard: a disallowed address, or, for HTTP requests, a scheme other than http/https, a port other than 80/443, or credentials in the URL.
unreachable_target424The destination you named did not resolve or refused the connection.
upstream_timeout424The destination you named did not answer before the declared timeout for this endpoint expired.
unsupported_content_type415The destination returned a type this endpoint does not handle. The source is not a type pdf-to-markdown reads; it reads PDF.
too_large413The response exceeded the size cap while streaming, or decompressed at more than the allowed ratio.
unprocessable422The bytes arrived but could not be processed, for instance a malformed document.
internal500Our fault, and in almost every case nothing is charged. The exceptions come after the call has run and the settlement call was made: the message "The work was done but settlement could not be completed", and the messages starting with "Payment settlement returned". There, whether the payment moved is unknown to us as well. The payment moves at most once: its nonce is single-use and settlement is idempotent, so a retry with it is never charged twice. A retry is normally refused with a 402: replay_consumed, an expiry reason, or a reason saying the nonce was already used once the transfer is on-chain. Only if our server restarted before the transfer was final can a retry run the call again. Even then, the payment moves at most once. To see whether it was settled, read the EIP-3009 authorizationState of its nonce on-chain; to call again, sign a new payment.
rate_limited429Too many calls from the same payer, or too many calls towards the same destination domain. Nothing is charged.
concurrency_limit,
registry_at_capacity,
upstream_unavailable
503We could not serve the call right now. Retryable: wait the number of seconds in the retry-after header, then repeat the call. concurrency_limit and registry_at_capacity mean we were at capacity: the call did not run. You can send the same payment again if its validBefore is still far enough away to cover another call. If it is not, the retry gets a 402 authorization_validity_too_short: sign a new payment. upstream_unavailable means our internal service failed or did not answer in time: the call may already have run and the payment may have been settled. A retry with the same payment can get a 402 whose error field says why. authorization_validity_too_short: too little time is left before validBefore to cover another call. This check comes before the replay check, so it does not tell whether the first payment was settled. replay_consumed (the payment was used and may have been settled) or replay_exhausted (too many attempts): sign a new payment. replay_in_flight: the first call is still running, so wait and do not pay again. invalid_exact_evm_nonce_already_used: the transfer is already on-chain, so this payment cannot move again. Sign a new payment. An authorization can be settled only once, since its nonce is single-use. To see whether it was settled, read the EIP-3009 authorizationState of its nonce on-chain.
misconfigured503The service is not configured to serve calls. The call did not run and nothing is charged. It needs a fix on our side, so there is no retry delay to wait for.

Input that is rejected before the tool runs (malformed JSON, schema mismatch, body over the size limit) is never charged.

Conversions

Ready-made calls to this API, one page each.

ConversionRecorded sampleSourceOutput
PDF to Markdowndocuments/markdown-guide.pdf1.1 KB249 bytes

Questions

How much does a call to the PDF to Markdown API cost?

$0.005 per call, paid in USD Coin on Base via x402. There is no account and no API key: the price is in the 402 answer, before anything is paid.

What is the timeout of pdf-to-markdown?

The endpoint itself runs for at most 40 s. max_timeout_seconds in the 402 answer is 61: that covers the whole call, including payment verification and settlement around it.

How large can a response from pdf-to-markdown be?

The response-size cap is 25 MB (26,214,400 bytes).

Is the pdf-to-markdown input schema free to read?

Yes. GET https://grist.tools/v1/pdf-to-markdown/schema returns the input schema for free, paid endpoints included.

What happens if a call to pdf-to-markdown fails?

The payment is verified before the handler runs and settled only after it succeeds, so a call refused by the gate, or one that fails inside the handler, is never settled. The single case that cannot be stated in advance is a settlement call that fails without answering: the error message says so when that is what happened.

What does pdf-to-markdown return?

A JSON object with 6 top-level fields: source_url, total_pages, extracted_pages, truncated, chars and markdown. The example on this page is a real call over documents/markdown-guide.pdf.