Inspiration

A small museum can hold extraordinary stories and still have a tiny team. Research, translation, image preparation, narration and accessible publishing usually mean coordinating several specialists, budgets and handoffs. I wanted a curator to commission the outcome and still understand every agreement and decision along the way.

Guild starts with Harbor Street, an imagined maritime museum built around twelve real archival photographs. The outcome is tangible: a beautiful exhibition that more visitors can read, in more languages, with the original sources always within reach.

What it does

Guild is a commission studio for coordinated AI work. The public Harbor Street reader has twelve archival photographs, eight chapters, six language choices, Arabic right-to-left reading, source notes and reading controls. It is clearly labelled as a curated Sample exhibition.

Behind that reader, the studio connects a commission, provider cards, the production floor, Passport, contract review, settlement and evidence history. Each package states its provider, scope, price and acceptance criteria. Passport constrains purpose, budget and expiry. The review room keeps a deliverable, its original agreement and its checks together so the curator can make an informed acceptance decision.

The provider market includes seven controlled addressed services, bilateral owner policies, a tested A2A JSON-RPC profile and three bounded MCP tools. These are controlled services, not seven outside businesses. The production floor uses working calendars and dependencies to explain what can finish and why.

Guild keeps generated work, automated checks, curator acceptance and provider payment as separate states. A draft does not become accepted because a model likes it; an operation does not become paid because an API acknowledged a request.

How we built it

The application uses Next.js, React and TypeScript on Vercel, with PostgreSQL for durable commands, commissions, jobs, authority, artifacts and financial state. Private Blob storage keeps media tied to immutable identities and independently checked bytes. Sharp and FFmpeg measure the actual image and audio files.

OpenAI handles structured research, localization, image descriptions, planning and original speech production. Google Gemini provides independent language verification. Actual saved research, image derivatives and six-language translation files are retained. Producers and independent checks have separate records, and an uncertain original request is preserved rather than silently submitted again.

Bryntum Gantt and Scheduler share a native scheduling project for calendars, dependencies, capacity and recovery proposals. PayPal Orders and the pinned APIMatic-generated TypeScript SDK support the guarded authorization/capture/void path; typed REST covers Payouts where the pinned SDK has no controller. PayPal Agent Toolkit is constrained to the qualified get_order read path. Human approval and server-derived amounts remain the financial boundary.

Challenges we ran into

The hardest problem was preserving the meaning of a result after a timeout, contract change or partial failure. A new screen must not quietly renew an old authority, replace an unknown request or attach a passing check to a different artifact. We built durable operation identities, immutable original receipts, exact artifact and contract hashes, and explicit recovery paths.

Multilingual production also taught us to separate measurable properties from human judgment. File format, duration and source linkage can be checked mechanically. Fluent language, historical interpretation and listening quality still need a person. The interface makes that distinction visible.

Accomplishments that we're proud of

  • The exhibition makes the outcome understandable immediately, with a strong maritime visual identity, six locales, Arabic RTL and inspectable archive sources.
  • Actual source-linked research, image derivatives and translation files are retained with their original production and verification records.
  • The original narration follow-up passes all 78 deterministic file checks. Independent recognition meets its advisory target for 60 of 78, and the other results remain visible. Human listening and fluent acceptance are still open.
  • Actual PayPal sandbox evidence records a $640 authorization followed by a confirmed zero-capture void, plus a genuine model-selected Agent Toolkit get_order call. Captures and provider payouts remain $0; a completed six-provider settlement is not claimed.
  • The deployed release passed its bounded hosted and native review, with original rows and session expiries preserved.

What we learned

Useful agent systems need agreements and evidence as much as generation. An ambitious result becomes manageable when each contribution has a clear scope and an inspectable history. The human should be able to see what was promised, what was produced, what was checked and what still needs a decision.

What's next

Complete the genuine accepted six-package commission and its independently confirmed sandbox settlement, expand provider onboarding, qualify the remaining accessibility and language-review journeys, and run studies with real curators and providers. Local work on brief onboarding, recourse, Gemini speech, AP2 exports and saved wallets is preserved separately and is not claimed as part of the current canonical release.

Try it

Start at https://paypal-ai-guild.vercel.app/exhibitions/harbor-street. Enter the waterfront, switch to Arabic, move between chapters and open the original archive source. The 27-second video shows actual interaction with this Sample reader.

Then open https://paypal-ai-guild.vercel.app and use its explicit sample entry to inspect the commission studio, Passport, package contracts and evidence. No personal PayPal credentials are needed to explore the sample. Sample acceptance does not send money. The README and capability ledger distinguish actual saved production, controlled fixtures and remaining gates. No real museum adoption, fluent acceptance or completed provider settlement is claimed.

Built With

Share this project:

Updates

Submission history