Public Beta

Playground

connecting…

Same product · any language · one fingerprint

Four real products in twelve languages — watch the same identity emerge, live.

Four real products, four different setup models — a brochure (folded sheet), business cards (flat sheet) and a booklet (multi-page) and a panel print (custom size on a rigid board) — each phrased in twelve languages. Pick a product, then a language, and watch the same identity emerge.

The interpretation (words → bindings) is pre-scripted per language — that is the AI agent's job in production. The term labels, envelope validation and the ptpge-cfs/1 fingerprint run live against this server. The hash is a deterministic function, not a model output.

ptpge_search_terms — free-text search

Type a buyer phrase in any of twelve languages — get the canonical term and its permanent URI.
Search is tokenized and language-agnostic — buyer phrases in any supported language hit the canonical terms. Try: ivory paper · black only · accordion fold · creasing · gold foil

ptpge_get_term — full context for one term

One URI in — full context out: labels in all languages, definition, place in the tree.
Tip: click a result id from a search, or try brochures · medium-paper · quantity

ptpge_get_version — dataset & server version

Which dataset and server version this playground is talking to.

ptpge_resolve_color_notation — the shorthand buyers actually write

"4/1" is a count of separations, not a colour. Watch the same notation resolve differently on different products — and ask when it must.

A buyer writes "CMYK on the front, black on the back" — or just 4/1. But "1" is black on one press and a named spot colour on another, so the number alone cannot identify the product. The resolver searches the combinations this product's envelope supports whose channel counts sum to that number.

Three outcomes, all correct: 4/4 resolves (4 is CMYK by trade convention). 4/1 on books resolves to black — that template supports no other single-channel ink, so there is nothing to ask about. 4/1 on business cards returns needs_disambiguation with a ready-made question, because black, white and spot are all possible. 4+1 asks about the separator first: in parts of the trade it means five inks on one surface, not front/back. The server never guesses — a guessed default would give two different products the same fingerprint. (SD-0071)

Test-prompt demo — the full chain, live

Envelope → validation → fingerprint on a fixed example order — the whole journey in three clicks.

"I want a brochure, 6 pages, A5 portrait, folded as a Z-fold, on ivory-toned heavy paper, full colour on one side and black only on the other, 500 copies."

The interpretation (words → bindings) is pre-scripted here — that is the AI agent's job. The lookups, envelope validation (incl. 250 gsm → weight band and the creasing advisory) and the ptpge-cfs/1 fingerprint run live against this server, exactly as an agent would call them.

Your own prompt — see what resolves

Describe a print job in your own words and see which pieces map to the specification.
Two ways to solve the same prompt. Agent mode (the green box — the primary path): a real model is handed the tools and runs the loop itself, from the outside, like any AI agent. Rules (the outlined buttons): the page searches 1–3 word n-grams and assembles them by rule — no model, deterministic, an approximation of what an agent does. Both end in the same place — a validated, complete configuration and one ptpge-cfs/1 fingerprint — or in the specification's question (SD-0083: no identity until the required axes are known). Both log every call to the transcript.
Agent mode — a real model runs the loop Complete prompt → one fingerprint. Leave out the fold or the orientation and the agent asks instead of guessing.
The model is handed the same tools an AI agent gets from tools/list and runs the loop itself: search terms, read the envelope, resolve colour counts, validate, fingerprint — every model turn and every tools/call lands in the transcript. Sponsored runs go through /agent/messages on this server — PTPGE pays, the key never leaves the server, prompts are not logged. Your own key is kept only in this browser's localStorage and sent straight to the provider.
For everyday use, switch on the connector in Claude or ChatGPT (https://mcp.ptpge.org/mcp) — this page is for watching the loop.
Session transcript — every MCP call this page has made clear
No calls yet. Every button logs its request and response here, chronologically — the conversation an AI agent would have with this server.

Every button above sends a real MCP tools/call (JSON-RPC over Streamable HTTP) to /mcp on this server — exactly what an AI agent does.