Tools

Returns the authoritative current time in one or more timezones, with UTC offset, ISO week number, day of year, weekend flag, and daylight-saving state. Use this instead of guessing the date: a language model has no clock, and its training cutoff is not today.

Converts one instant into any number of timezones, handling daylight saving transitions correctly. Naive inputs (no Z or offset) are read as wall-clock time in `from`; inputs carrying an offset are absolute.

Computes the elapsed time between two instants: exact totals in each unit, a calendar breakdown (years/months/days rather than 'thirteen and a half months'), the count of weekdays in between, and a human phrase. Negative when `to` precedes `from`.

Answers the questions a deadline actually poses: is this instant inside the window, how long until it opens or closes, and if it is outside, by how much. Optionally compares two windows for overlap. A window comparison that silently succeeds against the wrong bounds is a failure mode that does not raise an error anywhere — the ship window is missed, or the ASN is late, and the only symptom is a chargeback later. This computes the boundary instead of assuming it.

Parses a standard 5-field cron expression (or 6-field with leading seconds, or a macro like @daily), validates it, describes it in plain English, and computes the next run times in a given timezone with daylight-saving handled. Use it before writing a schedule to config — cron's day-of-month/day-of-week OR-semantics and step syntax are a common source of silently wrong schedules.

Runs a JSONPath query and returns the matching values with their concrete paths. Supports property access, array indexing (including negative indices), slices, wildcards, recursive descent (`..`) and filters (`[?(@.price > 10)]`). Pull three fields out of a large API response without carrying the whole document through context.

Takes JSON that almost parses and makes it parse: markdown code fences, prose wrapped around the object, trailing commas, single or smart quotes, unquoted keys, Python True/False/None, comments, NaN, and brackets left open by a truncated response. Every change is reported, so the caller learns what its generator got wrong rather than silently depending on a fixer. When the input cannot be salvaged it returns the precise line, column and a caret pointing at the offending character — which is what makes the next attempt succeed instead of guessing.

Converts a document to markdown or plain text: pass a public URL or the file itself as base64, and get back the content with headings, tables and lists preserved, at a fraction of the tokens that rendered pages cost. Use it when a harness has no native reader for the format — .docx, .xlsx, .odt and .numbers rarely have one — when a document is only a URL away, or when a long PDF's text matters and its layout does not. Handles PDF (.pdf), Word (.docx), Excel (.xlsx, .xlsm, .xlsb, .xls), OpenDocument (.odt, .ods), Apple Numbers, CSV, HTML, XML, and plain-text formats such as .txt and .md. The format is detected from magic bytes, not trusted from the file name, so a PDF served from a .php URL still converts. Two honest limits: a scanned PDF with no text layer has nothing to extract (this is conversion, not OCR), and legacy binary .doc and .ppt files are not readable — resave them as .docx or .pptx. Images are refused rather than described. Documents up to 10 MB.

Computes cryptographic digests and HMAC signatures. Returns hex and base64 at once, so a webhook signature can be compared in whichever encoding the provider uses. MD5 and CRC32 are included for checksums and legacy fixtures; they are not secure for authentication.

Evaluates arithmetic with correct operator precedence, parentheses, exponentiation, factorials, named variables and 30+ functions. Language models perform arithmetic by pattern-matching and quietly get long multiplications and compounding wrong; this computes the number instead of predicting it.

Resolves a URL and reports what actually happens to it — final status, the full redirect chain, content type and page title. Link rot is faster than it feels: one agent measured seven dead URLs out of 286 across fourteen days. This handles the three cases a naive checker gets wrong: an arXiv v1 to v2 redirect is a version change, not a break; a 403 is a blocked crawler, not a dead page; and a DOI can return 200 while the page behind it is now a retraction notice, which is the worst case because a simple status check passes and the reader gets something else. Response bodies are never returned — only status, headers, title and flags.

The difference between an agent believing it succeeded and knowing it did. http.check tells you a URL is alive; this tells you it is correct. Assert the status, response headers, values at JSONPath expressions, substrings in the body and a latency ceiling, and get back one boolean plus a per-check breakdown showing expected against actual. An agent that has just deployed, migrated or reconfigured something can prove the outcome rather than reporting the absence of an error, which is the usual way agents claim a success they have not actually achieved.

Answers the questions WHOIS used to answer, over RDAP — the structured protocol the registries themselves now serve — with no key and no scraping: who registered a domain, when it expires, which registrar and nameservers it uses, whether transfers are locked, and whether DNSSEC is signed. Give it an IP address instead and it returns the owning network: name, CIDR range, and the responsible registry. A domain that does not resolve in RDAP is reported as likely available to register — an answer, not an error. These are live registry facts that no training corpus can supply: registrations, expiries and transfers happen daily, and an agent that guesses at them guesses wrong.

Reads a raw ASC X12 interchange and returns it as JSON: delimiters taken from the ISA header, the ISA/GS/ST envelope decoded, every segment split into named elements, and a document-specific summary - purchase order lines for an 850, the HL shipment hierarchy and SSCCs for an 856, invoice totals for an 810. EDI is positional and delimiter-sensitive, and unreadable without a spec table; reading it by inspection produces confident nonsense, which is why this is a tool rather than a prompt.

Runs the structural checks a trading partner runs before rejecting a file: control numbers matching between the ISA/IEA, GS/GE and ST/SE header and trailer pairs, declared counts matching actual counts, and - for an 856 - the HL hierarchy. HL faults are the usual cause of ASN rejections and chargebacks: a parent ID naming no existing node, a duplicate HL01, an unknown level code, a missing carton level. Every finding names the segment and states what the value should have been.

Compiles structured JSON into standards-valid X12. The parts that get documents rejected are exactly the parts a language model cannot hold in its head: the 856 hierarchy needs sequential HL IDs where every level names its parent's ID, and on a multi-carton, multi-SKU shipment that bookkeeping goes wrong silently. The ISA header is fixed-width — 106 characters exactly, every element space-padded to size — and one character out crashes the receiver's translator. Segment counts in SE01, CTT01, GE01 and IEA01 must agree with what was actually emitted. All of that is computed here rather than written by hand.

Searches every tool by name, summary, description and keywords, and returns the closest matches with their endpoints and parameters. Use this instead of loading the whole catalogue: describe the job ("check whether a shipment is late", "fix broken JSON", "validate a barcode check digit") and call what comes back. Each result says why it matched, so a wrong match is obvious rather than plausible.

Returns everything needed to call a tool correctly: its complete JSON Schema, every parameter with type and default, and examples known to work. Pair it with tool.search — search to find the name, describe to learn the shape, then call. This exists so the catalogue does not have to be loaded into context up front.

Dispatches to any tool in the catalogue. Most clients cap how many tools they will hold at once — Cursor drops everything past roughly forty across all servers combined — so only a core set is listed by default. Everything else is reachable here. Find a name with tool.search, check its shape with tool.describe, then call it through this.