Sierra shipped Poppy 1.0 on November 13 — a machine-readable protocol that tells a customer’s AI agent exactly what it can do on a business’s website, without scraping, guessing, or simulating human clicks. Shopify switched it on for every merchant on the platform automatically. Meta’s Muse personal agent now looks for Poppy first on every site it visits. The Next Web reported that 100+ companies have joined as design partners, including Delta, Cisco, Intuit, Salesforce, Twilio, and Zoom. If you build anything that serves customers — or anything that acts on behalf of a customer — this is infrastructure you need to understand.
The Problem: Agents Are Still Screen Scrapers #
Right now, most personal agents interact with business websites the same way screen scrapers did twenty years ago: simulate a browser, navigate through pages, parse HTML, and hope the site’s structure hasn’t changed. This approach is brittle by design. It breaks on updates. It creates legal ambiguity around consent and impersonation. And it produces a poor customer experience because the agent is guessing, not knowing.
Poppy replaces guessing with declaration. A company publishes a file that says: here is who is authorized to represent customers, here is what an agent can read, here is what an agent can change. The agent doesn’t need to simulate anything.
How Poppy Works: One File, Existing Standards #
Every Poppy-enabled company publishes a JSON document at /.well-known/poppy.json. The file points an agent to the company’s OAuth server, defines what scopes are available, and lists the APIs or MCP servers the agent can use. The full specification is available at personalagentprotocol.org. A minimal implementation looks like this:
{
"protocol_version": "1.0",
"organization": {
"name": "My Store",
"domain": "mystore.com"
},
"auth": {
"issuer": "https://auth.mystore.com",
"sign_in": [{ "type": "direct", "scopes": ["poppy:read"] }]
},
"apis": [
{
"type": "openapi",
"url": "https://api.mystore.com/openapi.json",
"description": "Order management"
}
]
}
The scope system is intentionally minimal: poppy:read for viewing data (orders, balances, bookings), poppy:write for making changes. Custom scopes can be layered on top. The authentication runs on OAuth and DPoP — standards every backend team already knows. No new frameworks, no proprietary SDKs.
The Shopify Effect #
The most significant move in Sierra’s announcement is what Shopify did: automatically published a poppy.json for every merchant on the platform. Millions of stores, zero merchant effort. This is the network effect moment for Poppy.
Personal agent developers — building on Muse, Claude, or anything else — will implement Poppy support because Shopify is there. Other retailers will implement it to remain discoverable alongside Shopify merchants. Sierra effectively seeded the entire e-commerce ecosystem at launch, which is an unusually smart protocol adoption strategy.
Who’s Live Today #
Three financial and travel companies went live the same day as the 1.0 release:
- Mercury — Personal agents can check balances, categorize transactions, and initiate transfers
- Synchrony — Agents can access financing options during shopping flows
- Turkish Airlines — Agents can check fares, pull up bookings, and find Miles&Smiles upgrade availability
These aren’t demos. A personal agent with the right scopes can initiate a bank transfer or rebook a flight today, with explicit company consent baked into the protocol.
Poppy’s Place in the Protocol Stack #
Poppy is not competing with MCP. It sits above it. MCP handles capabilities — once an agent knows what tools to call and has permission to call them, MCP is how it calls them. Poppy handles the layer before that: discovery and consent. The Poppy spec site explicitly positions it as complementary to A2A, which covers agent-to-agent coordination.
It also complements x402, the AI agent payments protocol we covered earlier this month. Poppy establishes what the agent is authorized to do; x402 handles how it pays. Neither is a replacement for the other, and payments are deliberately deferred from Poppy 1.0.
What 1.0 Doesn’t Include #
Payments. The 1.0 spec deliberately excludes agent-initiated payment transactions. Sierra is working with Stripe and other payment providers on a future extension. For now, agents can read, search, and prepare — they cannot complete a purchase. This is a real limitation for shopping use cases, but a reasonable call for a protocol that’s trying to establish broad adoption before adding the most complex piece.
Publish Your poppy.json Now #
If you run a consumer-facing web service, publishing a poppy.json is a low-effort move with clear upside: your service becomes discoverable to the personal agents Meta, Anthropic, and others are deploying. The implementation cost is a JSON file and an OAuth scoping pass on your existing API. If your OAuth is already clean, this is a half-day project.
The official implementation kit and validator are available at poppy.md. If you already run an MCP server, listing it in your poppy.json is the fastest path to compatibility — agents using Muse or any Poppy-aware framework will find you automatically.