Hey HN, I'm the author.

Some background: I've been working on tooling for UCP (Universal Commerce Protocol) for a few months. UCP is the open standard Google and Shopify announced in January - it lets AI agents discover stores via /.well-known/ucp and complete purchases through standardized APIs.

I built Agorio because when I tried to build a shopping agent against the UCP spec, I had to:

1. Write my own profile parser that handles both capability formats in the spec 2. Build a checkout state machine (incomplete → requires_escalation → ready_for_complete → completed) 3. Create a mock merchant from scratch just to test against 4. Wire up LLM function calling with JSON Schema tool definitions

None of this was commerce-specific - it was all protocol plumbing. So I extracted it into a reusable SDK.

The key abstractions:

- LlmAdapter - two methods: chat(messages, tools) and modelName. Any LLM with function calling works. The Gemini adapter is ~100 lines. - ShoppingAgent - takes an LlmAdapter, runs plan-act-observe with 12 tools. Manages cart state, checkout sessions, order history. - UcpClient - fetches /.well-known/ucp, normalizes capabilities, resolves REST/MCP/A2A transports. - MockMerchant - full Express server with UCP profile, OpenAPI schema, 10 products, checkout flow, order tracking. Supports chaos testing with configurable latency and error rates.

Technical choices I'd like feedback on:

- Is a plan-act-observe loop the right pattern, or should I support ReAct / tree-of-thought? - Currently UCP-only. ACP client is planned for v0.2. Should I prioritize that? - The LlmAdapter interface is deliberately minimal. Too minimal?

Would love feedback from anyone building with LLM function calling or commerce APIs. Happy to talk UCP/ACP protocol details.