ACP is the open contract between an agent (the planner) and a commerce backend (the executor). Three layers, one shape: A2A capabilities declare what an agent can do; MCP tools expose it to LLM clients; A2UI events stream agent state to a rendering surface. The same protocol powers Eagle AI for Your Grocer, the multi-persona demos in this library, and any third-party agent client integrating with Delectable.
Most "agentic commerce" today is a thin wrapper: a prompt + an API + some retries. That breaks when the agent client changes, when the model changes, or when a new surface (kiosk, voice, in‑aisle assistant) needs the same intelligence. ACP separates the planning layer from the executing layer with a stable contract on both sides — so the agent can be Claude today, Gemini tomorrow, and the OpenClaw open-source agent next month, all hitting the same backend without the grocer rebuilding integrations.
Declarative capability registry. An agent advertises what it can do; clients dispatch by capability ID, not by URL. Capabilities have typed parameters and response schemas that any A2A-compatible client can validate against.
The same capabilities exposed as MCP tools. Drop-in usable by Claude, OpenAI, or any FastMCP-compatible client. Tool definitions auto-generate from the A2A capability registry — no duplication, no drift between protocol surfaces.
Server-sent event stream that lets a UI render agent state in real time: tool calls, intermediate results, drift signals, citations. The chat experiences in this library all consume A2UI — same event types, different render targets (web, mobile, kiosk).
The current COMMERCE_CAPABILITIES set in production. Any A2A-compatible agent can invoke these without understanding the underlying architecture (BigQuery, Memgraph, Redis, model routing) — the protocol abstracts it away.
Switch from Gemini to Claude to GPT-5 by changing one config. The protocol layer is stable; the model is interchangeable. Hedge against single-vendor risk.
Web, mobile, in-store kiosk, voice, in-aisle scan all consume the same A2UI event stream. Build the agent once; render anywhere the shopper is.
Third parties (CPG brands, ad networks, social platforms) integrate via A2A capabilities without exposing internals. The Ads and Social modules ride the same protocol; future modules will too.
A2A is an emerging open spec; MCP is from Anthropic and growing fast. Building on these means external agents (OpenClaw, custom planners) can integrate with us without bespoke work.
When Google announced the Universal Commerce Protocol (UCP) in October 2025, the bar moved: Search agents will start calling retailers directly with typed A2A requests, and Merchant Center adds a new conversational‑attribute feed. Delectable's UCP gateway is the on‑ramp — a curated, externally‑safe subset of our A2A surface, deployed in the grocer's own GCP project so shopper data stays put and the GCP consumption bills to the retailer (not us).
A Cloud Run service that fronts the grocer's commerce stack. Discovery at /.well-known/ucp; dispatch at /ucp/v1/agent. Speaks the UCP request/response envelope; routes intents to the A2A capability registry above.
Terraform-provisioned BigQuery dataset (products, inventory, prices, embeddings) plus Cloud SQL + pgvector for private RAG. Real-time inventory and loyalty data never leave the customer's VPC.
PIM → Merchant Center conversational-attribute transform. Emits dietary attributes, meal contexts, shopping missions, substitutability classes — the new fields Google's Search agent expects.
{
"agent": { "name": "google-search", "version": "1.0" },
"intent": "build_cart",
"parameters": {
"ingredients": ["oat milk", "rolled oats", "blueberries"],
"household_size": 2
}
}
{
"grocer": "giant-eagle",
"tenant_env": "prod",
"intelligence_source": "delectable-saas:food-graph",
"duration_ms": 184
}
{
"items": [
{"sku": "GE-0012", "name": "Oat milk, 32oz", ...},
{"sku": "GE-0034", "name": "Rolled oats, family pack", ...},
{"sku": "GE-0099", "name": "Frozen blueberries, 12oz", ...}
],
"subtotal": 21.97
}
Cross-tenant demo + the FSR-facing one-pager live in commerce/examples/cross_tenant_demo/ and commerce/intake/google-retail-pitch/.
The HyperGraph is the data moat. ACP is the integration moat. A grocer plugged into ACP today gets every future agent client, every future channel, every future LLM — without rewiring. That's the difference between an LLM wrapper and an infrastructure layer.