A summary of a web page (url) or your own text (up to 12,000 characters) as short, medium or bullets, in the language you name, with key points. Llama 3.3 70B. A page is read as in /api/read, robots.txt respected.

Current TLS certificates for a domain from Certificate Transparency logs (via CertSpotter): issuer, validity dates and days left, DNS names (up to 100 per certificate), and the distinct subdomains seen. Keyless tier: one request per call; answers are cached an hour.

The latest releases of a GitHub repository: tag, name, date, pre-release flag, the first 1,500 characters of the notes, and the number of assets. Needs the GITHUB_TOKEN secret on the deployment; without it the call answers 503 and is not charged.

Native balance and the major tokens (USDC, USDT, EURC, DAI, WETH, cbBTC and WBTC where listed) of an EVM address on Base (default), Ethereum, Arbitrum, Optimism, Polygon, or chain=all, with USD values from Coinbase spot prices. One Multicall3 call per chain. Zero balances are omitted, and only the listed tokens are checked, not every token the wallet holds.

Structured data from a web page (url) or your own text: give a schema of up to 20 field names with what each holds, and get a JSON object with exactly those fields, null where the text does not state one. Llama 3.3 70B. Text is capped at 12,000 characters; a model answer that is not JSON is retried once, then not charged.

ERC-721 facts for one token: owner, collection name and symbol, tokenURI, and its metadata (name, description, image) when the URI is http(s) or ipfs. The metadata fetch has a 5-second timeout and a 256 KB cap; a failed fetch is reported in metadataNote and does not fail the call.

The latest Form 4 insider transactions of a US company (ticker, CIK or name): insider name and role, transaction date, code (P buy, S sell, A grant, M exercise, F tax withholding), shares, price and shares owned afterwards, with the filing link. Common-stock lines only; derivative grants are not listed.

Translation with Llama 3.1 8B: up to 6,000 characters per call, formatting, names, numbers and URLs kept, source language detected when not given. Language codes like en, tr, de, fr, es, ar, zh, or language names.

ERC-20 transfers into and out of an address in a window of recent blocks (default 2,000, max 10,000), from Transfer logs, newest first, max 100 rows with token symbols. token limits the search to one contract. A busy address can be truncated: narrow the window to see the rest.

Reads the text of an image (JPEG, PNG, GIF or WEBP, up to 4 MB) with Llama 3.2 11B Vision on Workers AI. Send an https url or base64 image, and optionally a language hint. Returns the full text with its line breaks, the lines, a language guess and the model that answered. An image with no readable text is a 422 and is not charged.

Safety check of a text or a chat turn with Llama Guard 3: safe or unsafe, and the violated categories (S1 to S14, named in plain words). Send text, or messages (up to 20 turns). Classifies only; it does not change or block anything.

Issues or pull requests of a GitHub repository, searched by words and state (open, closed or all), newest updated first: number, title, state, labels, author, comments, dates and link. Needs the GITHUB_TOKEN secret; without it the call answers 503 and is not charged.

Companies from the official registers of Norway, Finland, the Czech Republic and Poland. Search by name (Norway, Finland, Czech Republic) or look up one company by its number: Norway 9 digits, Finland Y-tunnus, Czech IČO, Poland KRS (Poland by number only). Answers give name, id, status, legal form, address, incorporation date and industry; Norway lookups by number also list managers. France is not included (see the PR).

Describes an image from an https URL or base64 (JPEG, PNG, GIF or WEBP, up to 4 MB): a short description, visible objects, whether it contains text, and the answer to an optional question. Uses Llama 3.2 11B Vision; the model's answers can be wrong.

Named entities in a text or a page: people, organisations, places, dates and money amounts (from the chat model), plus e-mail addresses, URLs and phone numbers (from patterns), each de-duplicated with a count. With url, the page's readable text is used (robots.txt is respected).

Third-party script and iframe hosts on a page, each classified (analytics, ads, tag manager, chat, session replay, social, consent, CDN), the cookies its first response sets (names and flags only, never values) and whether a known consent manager is present. Static HTML only: scripts added at run time are not seen.

Labels up to 20 texts with 2 to 20 labels you choose, with a confidence from 0 to 1 and a one-line reason per text. One label per text, or several with multi true. Uses the chat model with JSON output; labels are as good as the model's reading of them.

Runs a read-only SPARQL query on Wikidata and returns the rows as plain values (ASK returns true or false). Only SELECT and ASK are allowed; updates are refused. A LIMIT of 100 or less is added when missing. 8-second limit. Data is CC0.

Event logs of a contract over recent blocks (default 1,000, max 5,000), newest first, max 100: block, transaction hash, topics and data. topic0 takes an event signature such as Transfer(address,address,uint256) or a 32-byte hash. Public nodes refuse very large results: narrow the range or add topic0.

Replaces personal data in a text with placeholders: e-mail addresses, phone numbers, IBANs (mod 97 checked), payment card numbers (Luhn checked only), IPv4 addresses and, with names true, person names found by the chat model. Returns the redacted text and counts per type. Original values are never returned.