Generate an EPC QR code (SEPA Credit Transfer / « Girocode », EPC069-12) — the QR a banking app scans to pre-fill a transfer, the payment companion to an invoice. Body { name, iban, amount?, bic?, purpose?, remittance? | reference? (RF) }. Returns PNG (default), SVG, or the raw EPC text. IBAN + RF are validated; strict EPC encoding (order, lengths, ≤331 bytes, level M). EUR only. Not a payment initiation — just the encoded data.

Read a Factur-X / ZUGFeRD invoice: send PDF bytes or a JSON { url }. Extracts the embedded EN 16931 CII into normalised data — seller, buyer, lines, VAT, totals, payment, dates, references — with the detected profile. The read side of e-invoicing (the FR mandate starts with reception). No embedded CII → { profile:"none", warning:"no_embedded_cii" }. ?attest=1 → Ed25519-signed.

Validate a Factur-X / ZUGFeRD invoice: send PDF bytes, JSON { url } or { pdfBase64 }, or raw CII { xml }. Returns a rule-based report — PDF/A-3 structure, EN 16931 CII anchors, arithmetic rules (BR-CO) and French PA « Flux 2 » checks (BR-FR) — with the detected profile. Structural + heuristic, not a certified conformance service. ?attest=1 → Ed25519-signed audit trail.

Generate a Factur-X invoice (PDF/A-3 + embedded EN 16931 CII XML) from structured data: send { invoice, profile }, get the PDF bytes (or JSON with base64 + conformity report via Accept: application/json). Targets EN 16931 and the French PA « Flux 2 » rules (payment mentions, business process, electronic addresses) ahead of the 2026-09 mandate. Technical artifact, not tax/legal advice; you remain responsible for data accuracy and PA transmission.

Live market price stats for a kind of x402 service, from the public Bazaar catalog. Pass ?category=vat or ?query=pdf extraction and get {sampleSize, priceUsd:{min,p25,median,max,p75}, networks, examples} in USD — is your price fair, what should you charge? Computed over a live sample of up to 1000 resources; <3 matches returns an honest 'insufficient market data'. Neutral: our own endpoints are never featured.

Verify on-chain that a transaction is the USDC payment you expected. Pass ?tx=0x… plus optional ?expectedPayTo=0x… and ?expectedAmount=0.01 (USD); ?network defaults to the deployment chain (base or base-sepolia). Decodes the tx's USDC Transfer logs and returns {verified, actualPayTo, actualAmount, asset, blockTime, confirmations, mismatch}. Unknown/unmined tx → 503 (unbilled). Add ?attest=1 for a signed receipt.

On-chain + config reliability report for any x402 seller, public data only, computed live. Pass ?payTo=0x… or ?url= (a third-party x402 endpoint) → {onchain:{firstSeenAt,totalIncomingTransfers,uniqueSenders,last30dInflows,note}, config, liveness, discovery} plus factual flags (new_seller, single_sender, config_issues, unreachable). On-chain counts all incoming USDC transfers, not only x402 settlements. No score, no verdict; fails open.

Extract text from a .docx (Word) document: send the docx bytes directly, or a JSON body with a public {url}. Returns the plain text, a best-effort markdown rendering (headings, bold/italic and lists; tables are flattened to paragraphs), and metadata (title, author). Text layer only — no OCR of embedded images. Max 10 MB.

Run up to 10 internal calls in one payment: POST {calls:[{path,input}]} and pay the SUM of the internal prices (≤ $0.05 total), executed in parallel, results returned in order. Batchable endpoints are the local, deterministic ones (validators, format transforms, time/locale) — no browser, network, composite or dynamic endpoints. Each result carries {path, status, body}; a bad request (unknown path, >10 calls, over cap) is a 400, not charged.

Row-level diff of two CSVs (a→b) — reconcile two exports. With `key` (a column name, or a column index when header:false), rows pair by key → added / removed / changed (per-field old→new). Without a key, a whole-row multiset diff gives added/removed only. Delimiter auto-detected (, ; tab |); tolerant RFC-4180 parsing. First row is the header unless header:false. Max 5 MB per input; reported buckets cap at 10000.

Rerank passages by relevance to a query via Cloudflare Workers AI (@cf/baai/bge-reranker-base). Send {query, documents[] (≤50, each ≤4000 chars), topK?}; returns [{index, score}] best-first — the missing link after embed/chunk, keeping the top chunks before a costly LLM call. Score is the model's relevance score (higher = better), named as it only compares within this model. Unavailable binding → 503, unbilled.

Verify C2PA / Content Credentials (provenance & tamper detection) in an image: send the image bytes or a JSON {url}. Returns Trusted/Valid/Invalid/None with per-manifest signature, trust-chain (issuer) and asset-hash checks plus C2PA actions. 'Trusted' only if the signer chains to the C2PA trust list, otherwise 'Valid'. Relevant to EU AI Act Art. 50. Verification only (embedded manifests) — not generation.

Extract a page's structured data: JSON-LD (schema.org), Open Graph tags, and best-effort microdata. Returns the parsed JSON-LD objects, detected schema.org types (Product, Offer, Review, Event, Recipe, Article…), microdata items and the og:* map. Send inline {html} or a public {url} (SSRF-checked). Malformed JSON-LD yields a {parseError} entry, never a crash. Microdata is best-effort; JSON-LD is the reliable path. Max 5 MB.

Mask PII in free text before you send it to an LLM or write it to a log. Pattern-based detection (not ML): emails, phone numbers (Google libphonenumber), IBANs (mod-97), payment cards (Luhn), URLs and IPs. Choose types[] and mode: mask (j***@d***.com, readable), remove, or label ([EMAIL_1], re-identifiable by index). Names and postal addresses are out of scope. The input text is never logged or stored.

Text → vector embeddings via Cloudflare Workers AI. Send {text} or {texts[]} (≤20, each ≤4000 chars); returns one float vector per input. Default model @cf/baai/bge-m3 (1024-dim, multilingual). The model is named in the response — embeddings only compare within the same model. Completes the RAG chain: readability → chunk → embed, no OpenAI account. Cosine-compare the vectors yourself. Unavailable model/binding → 503, not billed.

Extract values from a JSON document with a JSONPath expression or an RFC 6901 JSON Pointer, without returning the whole document. Send {data, path}; get {matches, count}. Path starting with $ = JSONPath ($.items[*].price, $..author, [0], [1:5], [*]); starting with / = JSON Pointer (/items/0/price). No-eval tree walk with bounded traversal; filter expressions [?(…)] are not supported by design.

De-duplicate a list of records or strings. Exact: group by a `key` field (dot-path) or, by default, the whole item (key-order-insensitive). Fuzzy (`fuzzy`: a field name, or true for scalar strings): group by normalized Levenshtein similarity ≥ `threshold` (default 0.85) — e.g. 'Jean Dupont' ≈ 'jean dupont' but 'Jean Martin' stays separate. Returns unique[] and duplicates[] (kept + dropped indices). Fuzzy ≤ 1000 items.

Validate a JSON value against a JSON Schema (draft 2020-12 default; 2019-09/7/4 via $schema). Send {data, schema}; returns {valid, errors:[{path, message, keyword}]} with a JSON Pointer path per failure. For agents checking structured/tool-call output before use. Runs on an eval-less validator (Ajv can't run on Workers). Remote $ref is refused (self-contained only); unsafe regex patterns are rejected before compilation.

Reshape JSON declaratively — predictable, not a language. `mapping` mirrors the output shape; its leaf strings are dot-paths into `data` (a.b.0.c; negative index ok; 'contacts.*.email' projects over an array; '=text' escapes a literal). Objects/arrays nest; number/boolean/null are literals. Missing paths → null. Field rename/flatten/projection to fit a payload to a target schema. Distinct from /v1/jq (declarative vs query language).

Test a regex against one or many inputs, safely. Send {pattern, flags?, input|inputs[], mode?}: 'match' returns the first match + capture groups, 'matchAll' every match, 'replace' the rewritten string (needs {replacement}). A mandatory ReDoS guard statically rejects catastrophic patterns (nested quantifiers, `.*.*$`, backrefs) BEFORE running — on Workers no timeout can stop a runaway regex. For agents that generate and dry-run regexes.