@agentpos-alexa/bridge
Integrates with Amazon's Alexa+ program to expose an MCP add-on for online stores, including UCP-compatible checkout sessions, MCP Apps, Amazon Wallet payment handlers, and merchant agents on Bedrock AgentCore.
Provides a payment rail for real USDC settlement on Stellar through the org.x402.stellar interface, using an open facilitator.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@@agentpos-alexa/bridgeBuy a dozen roses and pay with Stellar"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
AgentPOS for Alexa+
Turn any existing online store into an Alexa+ add-on with in-conversation checkout: MCP server, UCP checkout, merchant policies and verifiable receipts, with the store as merchant of record and no marketplace in between. The merchant chooses how to get paid.
Alexa+ for Builders is for Priceline. AgentPOS is for the corner store.
Status: in development for the Amazon "Build, Ship, Shape" Developer Hackathon 2026 (Alexa+ track, AWS Builder and Open Source mini challenges). Public from the first commit.
What this repository is
Amazon defines an Alexa+ add-on as "the bridge you deploy" between an MCP server and Alexa+.
This repository is that bridge for stores that run AgentPOS: it sits
in front of any AgentPOS store, consumes the store's public surfaces (MCP, REST API,
/.well-known/ucp, checkout) and exposes the full Alexa+ contract:
an MCP server (Streamable HTTP, spec 2025-11-25) with voice-ready item data and MCP Apps (product carousel, item card, order card, receipt card with a verified badge);
the UCP-compatible checkout Alexa+ documents:
/checkout-sessionscreate, get, update, complete and cancel, OAuth 2.0 bearer auth, 24-hour idempotency, 6-hour sessions, business errors asmessages[];payment rails behind one interface, chosen by the merchant, never by us:
the merchant's own PSP (Stripe, in test mode for the demo): the charge lands in the store's dashboard with the store as merchant of record;
Amazon Wallet (
com.amazon.payments.network_token,stored_payment_method): implemented against the published contract and exercised with a clearly labeled simulated PSP while the Alexa+ program is in preview;USDC on Stellar (
org.x402.stellar) when the agent brings a wallet: real settlement through an open x402 facilitator;
merchant-side agents on Amazon Bedrock AgentCore with Strands: onboarding (turns a URL into a voice-ready catalog and proposed policies; a human confirms), catalog (answers what a flat catalog cannot) and policy guardian (adds context to the store's deterministic policy and can hand the order to a human);
a simulated Alexa+ client (
apps/simulator), the official alternative path for the track: an Echo Show style web app, voice in and out, whose agentic client runs on Bedrock and drives the bridge end to end.
The store side (catalog, quotes, merchant policies, human approval queue, signed receipt
chain, agent identity, settlement, WooCommerce order creation only after settlement) is
AgentPOS itself. This bridge runs today against the public demo store at
https://demo.agentposhq.com and needs nothing else.
Related MCP server: Maple Cart MCP
Layout
packages/store-client typed client for the public surfaces of an AgentPOS store (discovery, REST, x402 challenge)
packages/fixture-store test double of an AgentPOS Store: the real OpenAPI shapes, a bakery with physical goods, recorded demo responses
packages/bridge the add-on: MCP server for Alexa+, MCP Apps views, UCP checkout sessions, payment rails, UCP profile
packages/agents onboarding, catalog and policy guardian agents (Strands on AgentCore, Bedrock)
apps/simulator simulated Alexa+ experience (Echo Show frame, host checkout pattern, inspection summary)
docs/ strategy, submission draft, architecture, Alexa+ MCP design rules, security, AWS integration, usage events, friction logRun
Requirements: Node >= 22.13 (uses node:sqlite unflagged, no native dependencies), pnpm 11+.
pnpm install
pnpm typecheck
pnpm test
pnpm dev:fixture-store # a bakery-shaped test double of an AgentPOS Store on :8790 (optional)
pnpm dev:bridge # registers AGENTPOS_STORE_URL (default: the public demo store) and listens on :8787
pnpm dev:simulator # the simulated Alexa+ experience on http://127.0.0.1:8788 (builds its web app first)To run the bridge against the bakery instead of the public demo store, set
AGENTPOS_STORE_URL=http://127.0.0.1:8790. The simulator needs the same
BRIDGE_BEARER_TOKEN as the bridge: set it in .env for both, or copy the one-run token the
bridge prints at boot when the variable is missing.
Environment variables are listed in .env.example; bash scripts/setup-wizard.sh walks a
human through the ones only a human can create (AWS keys for Bedrock, model access, the
Stripe test publishable key) and writes them to .env. Nothing in this repository ever holds
a merchant's or a household's private key or PSP secret.
curl http://127.0.0.1:8787/health
curl http://127.0.0.1:8787/stores/demo-agentposhq-com/.well-known/ucpThe bridge
The bridge is multi-tenant: every Store it serves lives under /stores/{slug}/, and the
first one is registered at boot from AGENTPOS_STORE_URL. Every response carries a
Request-Id that is the trace id across the bridge's JSON logs, the Store request and the
settlement.
MCP for Alexa+ at /stores/{slug}/mcp: Streamable HTTP, spec 2025-11-25, bearer auth.
Six tools, one intent each: search_items, get_item, get_policies, start_checkout,
get_order, get_receipt. Every result speaks first (a short text block) and carries
structured content with prices as integer minor units; every failure is a typed error, never
an empty result. Four tools carry an MCP Apps view (carousel, item card, order card, receipt
card) that the bridge serves as a ui:// resource: one self-contained HTML file each, built
from packages/bridge/apps/ with pnpm --filter @agentpos-alexa/bridge build:apps and
committed, so nothing needs building to run.
UCP checkout at /stores/{slug}/checkout-sessions (create, get, update, complete,
cancel), as Alexa+ documents it. Sessions are priced from the Store's catalog, quoted by the
Store once a delivery address exists, and completed only when a payment rail reports
settlement with a reference. Idempotency-Key is required on every state change and replayed
for 24 hours; every response is validated in tests against the vendored official UCP schemas
(packages/bridge/vendor/ucp). Platforms get bearer tokens from POST /oauth/token with
client credentials (BRIDGE_OAUTH_CLIENT_ID and BRIDGE_OAUTH_CLIENT_SECRET).
Usage events: every tool call, rail call and model call lands in usage_events
(docs/USAGE-EVENTS.md); pnpm --filter @agentpos-alexa/bridge usage:export writes the CSV
and --summary the totals behind docs/COSTS.md.
The simulator
An Echo Show frame with the four display modes Amazon documents (inline, fullscreen,
voice-only, hydrated), the Enabled add-ons the bridge serves, voice or text input, and an
inspection summary per turn (.data/inspection-summary.json) that checks what Amazon's
guide asks for: voice text first, three to five carousel items, first item within 500 ms,
typed errors. Until the Household agent lands, a scripted router with no model maps the text
to one Bridge tool, and the interface says so.
Checkout is the host's own pattern, as it is in Alexa+: when the customer confirms what to buy, the simulator opens a UCP checkout session on the bridge, fills the Demo household's synthetic address, shows the store's quote with the payment handlers the session accepts (the Amazon handlers, labeled SIMULATED), and completes with the chosen one. The order card and the receipt card that follow are the bridge's own MCP Apps views.


Principles we do not bend
The store is merchant of record. No commission, no custody, no marketplace.
The merchant chooses the rail. The bridge never holds a PSP key; the store executes the charge.
Nothing scanned is published without human confirmation.
No order is created without settled payment and a settlement reference.
Every protocol detail (MCP, UCP, x402) lives behind an adapter and is pinned to a version.
Amounts are integers in minor units. Never floats for money.
Cryptography only through official SDKs (
@x402/*,@stellar/stellar-sdk, MCP SDK).Simulated means labeled: in code, logs, responses and UI, and never switchable to a real PSP by configuration.
Strategy
Why the rails are ordered this way, why the simulator comes second and how impact is
measured: docs/STRATEGY.md.
AWS integration
See docs/AWS-INTEGRATION.md (services, why each, where in the code).
Security by design
See docs/SECURITY.md.
Friction log
Written while building, one entry per obstacle with Amazon or AWS tooling:
docs/FRICTION-LOG.md.
License
Apache-2.0. Copyright 2026 NovaCorpAI SpA.
This server cannot be deployed
Maintenance
Related MCP Connectors
Agent-native commerce with trusted catalog, durable carts, and Stripe Checkout via MCP and UCP.
Agent Commerce MCP — agent-native A2A storefront. Discovery, Stripe checkout, affiliate program.
Agent Commerce Protocol MCP — bridges Stripe ACP + Google AP2 + Coinbase x402 for agent payments
Hosted Amazon Seller Central and Amazon Ads MCP server for Claude, ChatGPT, Cursor, and agents.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables interaction with Amazon services through AI assistants, allowing users to search products, manage their cart, view order history, and place orders using natural language.6 npmMIT
- FlicenseNot gradedqualityDmaintenanceEnables e-commerce shopping assistant capabilities including product search, cart management, payment processing, and order fulfillment via MCP tools.-
- AlicenseAqualityDmaintenanceEnables AI agents to discover, subscribe to, and call paid agent services from the AgentStorefront marketplace directly within MCP-aware clients like Claude or Cursor.422 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables Alexa+ to manage your tasks and calendar through conversational commands, including adding, listing, completing tasks, and generating AI-powered daily plans with AWS Bedrock.MIT