AI agents that shop, book and bank for you have had one big problem: websites were built for humans, so agents had to guess.
On 13 November 2026, Sierra released Personal Agent Protocol 1.0, nicknamed Poppy, a shared rulebook for how your personal agent introduces itself to a company and what it is allowed to do there.
It arrived with real reach on day one. Shopify switched it on for every store, and Meta’s Muse agent now looks for Poppy first on every site it visits.
This guide explains what Poppy is in plain words, how the discovery file and sign-in work, what it means for shoppers, Shopify merchants and developers, how it compares with MCP, A2A, UCP and ACP, and what is still unknown.
- What: an open standard (Apache 2.0) that lets a personal AI agent find a company, identify itself and act within permissions you approve.
- Who made it: Sierra, the customer-service AI company led by Bret Taylor and Clay Bavor, first shown with Meta in October.
- Day-one reach: every Shopify store now publishes a Poppy file automatically, and Muse checks for Poppy first.
- How it works: a file at
/.well-known/poppy.json, a signed session header, and normal OAuth sign-in with scopes you pick. - Live examples: Mercury (balances, transfers under approval rules) and Turkish Airlines (fares, bookings, upgrades you approve).
- Catch: the spec text still says “draft 1.0” and allows breaking fixes. There is no documented opt-out for Shopify merchants yet.
What Poppy Is, in Plain Words
Picture a personal agent like Meta Muse or ChatGPT dots arriving at an online store. Today it has to scrape pages, click buttons and hope nothing changes. The store has no clean way to know which agent it is talking to, or which customer that agent works for.
Poppy gives both sides a handshake. The company publishes a small file that says “here is who we are, here is how to sign in, and here are the ways an agent can talk to us”. The agent then opens a session, and the customer approves exactly what it may do.
In Sierra’s words, Poppy “enables personal agents to interact with companies through business agents, websites, and APIs”. It mostly glues together existing standards: OAuth for sign-in, MCP and OpenAPI for APIs, and Google’s Universal Commerce Protocol (UCP) for shopping carts and checkout.
From Draft to 1.0: Timeline
Poppy moved fast. Sierra published four frozen drafts in five weeks while design partners tested them against real websites, APIs and agents.
| Date (2026) | Milestone | Source |
|---|---|---|
| 9 Oct | Draft 0.1 published; Sierra and Meta unveil Poppy with first partners (Meta, Genesys, Instinct, NiCE, Rocket, Shopify, Stripe, Walmart) | Sierra blog, poppy.md |
| 24 Oct | Draft 0.2 | poppy.md status page |
| 4 Nov | Draft 0.3 | poppy.md status page |
| 10 Nov | Draft 0.4, described as the 1.0 release candidate | poppy.md status page |
| 13 Nov | Version 1.0 released; Shopify enables it for all merchants; Muse turns it on for all users; 60 new design partners | Sierra blog, The Next Web, Unite.AI |
Version 1.0 added staged adoption (start small, add features later), industry extensions, and guidance for companies and agents to send each other feedback.
How Poppy Works Under the Hood
You do not need to read the 100,000-plus-character spec to understand the flow. It comes down to four steps.
- Discover. The agent fetches
https://{domain}/.well-known/poppy.jsonover HTTPS. The file lists the company, its sign-in server and its routes. - Start a session. Every request carries a signed
Poppy-Sessionheader that identifies the agent and a pseudonymous user ID. - Act as a guest. Simple questions such as stock levels or return rules can run without signing in.
- Sign in only when needed. For account tasks, the customer signs in with standard OAuth and picks the scopes, such as read-only order history.
The discovery file
Only two fields are required: protocol_version (currently “1.0”) and organization (name and domain). The file must also offer at least one route: agent (the company’s own AI agent), apis (OpenAPI or MCP), or web (a website browser session). Here is a trimmed example based on the spec:
{
"protocol_version": "1.0",
"organization": { "name": "Example Co", "domain": "example.com" },
"auth": {
"scopes": { "orders:read": "View orders and returns" },
"scope_hints": { "orders:read": { "changes": "none", "personal_data": true } }
},
"agent": { "protocols": [ { "type": "poppy",
"endpoint": "https://api.example.com/poppy/conversations",
"description": "Orders and account help" } ] },
"web": { "browser_session_endpoint": "https://example.com/poppy/browser-session" },
"apis": [ { "type": "openapi", "url": "https://api.example.com/openapi.json",
"description": "Orders" } ]
}
The scope_hints are a nice touch. Each scope says whether it makes changes (none, reversible or irreversible) and whether it touches personal data, so the agent can warn you before risky actions.
Key technical pieces
| Piece | What the spec says |
|---|---|
| Sign-in | OAuth authorization code with PKCE (preferred), device sign-in (RFC 8628), refresh tokens; PAR and DPoP optional |
| Agent identity | client_id is an HTTPS URL with the agent’s metadata; client assertions use private_key_jwt (ES256 and PS384 required), max 5-minute lifetime |
| Headers | Poppy-Session, Poppy-Version, Poppy-Extensions, plus a Bearer or DPoP token when signed in |
| User IDs | Pairwise and opaque: stable for one company, different across companies, no personal info |
| Delegation | Short-lived JWT (up to 24 hours) that vouches for the key signing session tokens |
| Extensions | Industry add-ons such as the commerce extension co-written by Shopify; custom ones must use your own domain name as a prefix |
Who Supports It Today
Sierra says more than 100 companies are now design partners. Being a design partner does not mean a company’s site already works with Poppy, so it helps to separate “live” from “on board”.
| Company | Status | What an agent can do |
|---|---|---|
| Shopify | Live for every merchant | Each store auto-publishes a Poppy file; carts and checkout run through UCP; agent must ask the shopper before certain actions |
| Meta Muse | Live for all Muse users | Looks for Poppy first on every site it visits |
| Mercury | Live | Check balances, review and categorise spending, manage budgets, start payments that still follow the account’s approval policy |
| Turkish Airlines | Live | Check fares, flight status and stopovers, find bookings by reference and surname, search Miles&Smiles business-class upgrades you approve before paying |
| Synchrony | Exploratory | Product discovery with financing and offers, potential reach of 500,000+ merchant locations |
| 60 new partners | Design partners | Includes Affirm, Delta Air Lines, Cisco, Intuit, Salesforce (Agentforce, Commerce Cloud), Stripe, Twilio, Wayfair, Wix, Zoom, Hyatt, SoFi, GoFundMe and more |
What It Means If You Use an AI Agent
If you use Muse, you get Poppy automatically and do not need to switch anything on. When Muse visits a Shopify store, it can now talk to the store through a proper channel instead of guessing at web pages.
The practical changes you should notice:
- Clear sign-in screens. When a task needs your account, you see the company’s own consent page. Opening a sign-in link must not approve it by itself.
- Scoped access. You choose read-only or write access. A company must not grant more scopes than the agent asked for.
- Approval for big actions. Purchases, upgrades and payments ask you first, following the company’s own rules.
- Human handoff. Conversations label whether a person or an AI wrote each message and support passing you to a human.
If you use ChatGPT, Claude or Gemini agents, Poppy support depends on those companies adopting it. None of them were named as Poppy implementers in the 1.0 announcement. For how OpenAI handles agent payments today, see our ChatGPT Wallet guide.
Shopify Merchants: Check Your Store’s Poppy File
Shopify says merchants need to do nothing. Every store now publishes a Poppy file, and the commerce extension links Poppy to UCP and to UCP Ask, so a shopper’s agent can talk directly to your store’s agent.
Still, it is worth seeing what agents see. Here is a quick check:
- Open a browser and go to
https://yourstore.com/.well-known/poppy.json(use your own domain or your myshopify.com address). - Confirm
organization.domainmatches your store’s domain. The spec requires this match. - Look at the routes listed under
agent,apisandwebto see which channels agents are offered. - Review your store’s return, shipping and stock information, because agents will answer shopper questions from what your store exposes.
- Check your Shopify admin’s AI and sales-channel settings for any new agent options as Shopify updates its help pages.
Developers: Add Poppy to Your Own Site
If you run your own site or app, Sierra’s implementation kit is built to be used with a coding agent. It ships as a skill named implement-personal-agent-protocol plus a command-line validator that acts as a test personal agent.
What you need
- A coding agent with access to the repositories you will change (anything that supports skills or can fetch a URL).
- Node 22.18 or later for the validator, which has no dependencies.
- A deployment the validator can reach over HTTPS, plus an HTTPS URL you control to host the test agent’s public files.
Install the kit
In Claude Code, run these two commands, then ask Claude to add Personal Agent Protocol support:
/plugin marketplace add personal-agent-protocol/personal-agent-protocol
/plugin install personal-agent-protocol@personal-agent-protocol
For other agents, download implement-personal-agent-protocol.zip from the kit page at personalagentprotocol.org/kit. Validator help is available with node <skill-dir>/validate/index.ts help.
The five stages
| Stage | What happens | Validator command |
|---|---|---|
| 1. Survey | Agent reads your code and config, records systems in plan/survey.md (GET requests only) |
survey |
| 2. Design | Agent maps the protocol to your systems in plan/design.md; you approve scope before any build |
none |
| 3. Build | One milestone at a time | check --milestone |
| 4. Verify | Full check against your deployment, then checklist.md for things the validator cannot test |
check --design plan/design.md |
| 5. Report | Drafts an implementation report; nothing is sent until you file it | report |
If your site already has an MCP server, you can list it as an apis route. Our guides to building an MCP server in Python and MCP security best practices cover that side.
Poppy vs MCP vs A2A vs UCP vs ACP
Agent standards are multiplying, and they overlap less than the acronyms suggest. Poppy is mostly about identity and permission at the company’s front door.
| Protocol | Backed by | Main job | Overlap with Poppy |
|---|---|---|---|
| Poppy | Sierra (Apache 2.0) | Personal agent finds a company, identifies itself, gets scoped access | n/a |
| MCP | Anthropic | Tools and data an AI app can call | Low; MCP can be a route inside Poppy |
| A2A | Google, now Linux Foundation | Agent-to-agent messaging | Some |
| UCP | Google with Shopify and retailers | Product discovery, carts, checkout, after-sale | Some; Poppy hands off to UCP for shopping |
| ACP | OpenAI and Stripe | In-agent checkout and payment tokens | Some |
| TAP | Visa and Cloudflare | Agent identity on each request | High |
| PACT | Decagon and Instinct | Agent identity and customer authority | High |
Comparison based on the poppy.md comparison page and Sierra’s 1.0 post.
For a wider look at MCP tools you can already use, see our list of the best MCP servers.
Privacy and Safety Rules Built In
The spec has several rules that matter for ordinary users, not just engineers:
- No cross-site tracking by ID. Your user ID differs for every company, so two companies cannot match you through Poppy.
- Prompt-injection guard. Content from companies is treated as data, not instructions, so a shady page should not be able to steer your agent.
- Secrets stay out of the model. Tokens, keys and authorisation codes must not enter the AI’s context.
- Controlled browsers only. A browser may join a session only if the agent controls it, never your own browser.
- No tokens in URLs or redirects. Session tokens must never appear in URLs or be sent to redirect targets.
The spec also flags real risks honestly, including sign-ins started by someone else on another device. Always check the company name and link host before approving a sign-in request.
Limits and Open Questions
- “1.0” is still a draft. The spec and kit say “draft 1.0”, and fixes may break compatibility.
- No merchant opt-out documented for Shopify’s automatic Poppy files.
- Payments are not fully covered yet. The roadmap lists payments, push notifications, attachments and self-hosted personal agents as future work.
- Big agents are missing. OpenAI, Anthropic, Google and Amazon were not named as Poppy agents in the release, and Amazon prefers shoppers use its own agent.
- Adoption is early. Only a handful of companies are confirmed live beyond Shopify.
Troubleshooting Common Errors
These are the error codes defined in the spec and what they usually mean when you test an implementation.
| Error | HTTP | Meaning and fix |
|---|---|---|
invalid_scope |
OAuth | The agent asked for a scope the company does not list. Match scopes to auth.scopes in poppy.json. |
invalid_session |
400 | The session JWT failed checks. Verify the delegation JWT, signature, aud, htm/htu claims and clock skew. |
session_ended |
400 | The company closed the session. Start a new session with a fresh sid. |
version_unsupported |
400 | Version mismatch. The agent should refetch poppy.json, pick a supported version and retry once. |
extension_required |
403 | The company needs an extension (named in the response) that the agent did not declare in Poppy-Extensions. |
| Rate limited | 429 | Wait at least the Retry-After time before retrying. |
| Discovery fails | n/a | poppy.json must be served over HTTPS at the exact path; redirects must end at that path with no query string, and organization.domain must match the host. |
Session errors use 400 instead of 401 on purpose, so OAuth libraries do not throw away valid tokens. If your client treats every 400 as fatal, special-case these codes.
My Take
Poppy’s smartest move is distribution, not technology. Most of it reuses OAuth, MCP, OpenAPI and UCP. But getting every Shopify store and every Muse user on day one gives it a head start that a standards committee rarely gets.
The privacy choices are good: pairwise user IDs, secrets kept out of the model, and company content treated as untrusted. Those are exactly the problems that have caused real agent incidents this year.
The weak spot is that “1.0” label on a spec that still calls itself a draft, plus no visible opt-out for merchants. If OpenAI, Google or Anthropic back a rival front-door standard such as TAP or PACT, businesses could end up supporting three of these.
For now, if you sell on Shopify, check your poppy.json. If you build agent-facing APIs, the kit is worth a weekend in a test environment. Our Muse phone calls guide shows how fast Meta is pushing its agent into everyday tasks.
FAQ
What is Poppy in AI?
Poppy is the nickname for Personal Agent Protocol, an open standard from Sierra that lets personal AI agents find a company, identify themselves and act with permissions the customer approves. Version 1.0 was released on 13 November 2026.
Is Poppy free and open source?
Yes. The spec and implementation kit are published under the Apache License 2.0 on GitHub at personal-agent-protocol/personal-agent-protocol.
Do Shopify merchants need to do anything?
No. Shopify says every store now publishes a Poppy file automatically. You can view yours at yourstore.com/.well-known/poppy.json.
Can I turn Poppy off on my Shopify store?
Neither Shopify nor Sierra has documented an opt-out as of 15 November 2026. Watch Shopify’s help centre for updates.
Which AI agents use Poppy?
Meta Muse supports it for all users and checks for Poppy first on every site. Sierra did not name ChatGPT, Claude or Gemini agents as supporters in the 1.0 release.
Is Poppy a replacement for MCP?
No. MCP describes tools an AI app can call. Poppy is about identity and permissions, and a company can list an MCP server as one of its Poppy routes.
Can an agent spend my money through Poppy?
Only within what you approve. Scopes are chosen at sign-in, and examples like Mercury payments and Turkish Airlines upgrades still require your approval. Full payment support is on the roadmap.
Does Poppy let companies track me across sites?
The spec uses pairwise user IDs that differ for each company, so they cannot link you through Poppy itself. Normal sign-in data you share with each company still applies.
Sources
- Sierra blog, “Releasing Personal Agent Protocol 1.0” by Bret Taylor and Clay Bavor (13 Nov 2026)
- Personal Agent Protocol 1.0 specification, personalagentprotocol.org/docs/1.0/spec
- Personal Agent Protocol company implementation kit, personalagentprotocol.org/kit/company
- Sierra on X, Poppy v1.0 announcement (13 Nov 2026)
- The Next Web, “Sierra releases Poppy 1.0 as Shopify and Meta’s Muse sign on” (13 Nov 2026)
- Unite.AI, “Sierra Releases Poppy 1.0 With Shopify and Meta Muse Support” (Nov 2026)
- poppy.md, spec status, versions and protocol comparisons (Nov 2026)
- Sierra blog, “Introducing Personal Agent Protocol” (9 Oct 2026)




