Arthera

Inspiration

AI agents can understand requests and propose actions, but commerce requires more than a convincing response. Orders need accurate prices, explicit approval, reliable payment verification, and a clear record of what happened.

We built Aria Commerce to explore how an AI assistant can help with these workflows while keeping people in control of financial decisions.

What it does

Aria Commerce is a commerce plugin for Aria Code, developed as part of Arthera's reusable Aria Plugins ecosystem.

A user can describe an order, review an itemised invoice, approve the exact amount, and proceed through checkout. Catalogue prices are validated by the server rather than accepted from model-generated text.

The Aria Code integration exposes three MCP tools:

Browse the product catalogue.
Create an order draft.
Read an order's status.

Approval and payment actions require human confirmation in the browser.

The deployed public demo currently uses a local request parser and simulated payments. A PayPal Sandbox adapter and AI Gateway adapter are implemented, but live integration testing is still pending credential configuration.

How we built it

We built the backend and MCP bridge in Go and deployed the service on Google Cloud Run.

Firestore stores orders, audit events, shared request quotas, and durable payment-operation records. Session and merchant credentials are stored in Google Secret Manager.

The HTTP service and recovery worker share the same commerce logic. A separate Cloud Run Job, triggered by Cloud Scheduler, checks pending operations and reconciles uncertain results. This avoids depending on a background process inside a web server.

The PayPal Sandbox adapter includes checkout, capture inspection, and webhook verification. The AI Gateway adapter supports model-assisted order drafting.

Challenges we ran into

One major challenge was separating an agent's ability to prepare an order from a person's authority to approve payment. We addressed this with distinct browser and agent credentials and server-enforced restrictions on approval and payment actions.

Another challenge was handling uncertain payment results. A provider may complete an operation even when the response is lost, or the application cannot save the result. We persist an operation record before contacting the provider and use stable request identifiers and transactional updates to support recovery.

Deploying on Cloud Run also required shared storage, persistent session secrets, and health-check paths compatible with the platform.

Accomplishments that we're proud of

Deployed a working public commerce demo on Cloud Run.
Verified the simulated order-to-receipt workflow.
Confirmed that orders survive a service revision replacement.
Connected the actual Go MCP bridge to the cloud service.
Verified that agent credentials cannot approve an order.
Successfully executed the recovery Job and authenticated scheduled triggers.
Passed automated Go checks and payment-failure tests.

What we learned

Reliable agentic commerce requires clear boundaries between intent, approval, execution, and verification.

We learned to treat model output as a proposal, use trusted catalogue data for pricing, persist operations before external calls, and distinguish a demonstrated feature from an integration that still needs live testing.

What's next for Arthera

Our next milestone is to configure and validate real AI drafting and PayPal Sandbox transactions, including checkout, capture, webhook delivery, and recovery from uncertain provider results.

We also plan to improve merchant authentication, operational monitoring, and the reusable plugin interfaces so additional Aria services can share the same foundations.