Poppy 0.3 vs 0.4 vs 1.0: Who Publishes It, and What Muse and Dot Really Use
Guide11 min readFacts checked on the date above
Poppy,tested.
0.3 vs 0.4 vs 1.0
Short answer: Poppy works today, but not the way the launch posts suggest. Muse speaks an
in-between dialect (0.4 headers, 0.3-style operations), uses a company's MCP APIs far more readily than its
chat agent, and ignored a file that offered only version 1.0. Only 6 of about 105 design partners publish a
poppy.json at all, none of them at 1.0. Shopify serves 0.4 for every store, Wix serves 1.0 for its sites. In
our tests Muse used Poppy end to end, up to a test order, on a store that exposed MCP tools on its own domain, and browsed
the pages of a Shopify store whose file only points to UCP. On a Wix store it used Wix's MCP tools, slowly and with
many approval prompts. Dot by OpenAI, also a design partner, never used Poppy: it read pages, and on a Shopify store it used Shopify's
in-page WebMCP tools instead.
What Poppy is, in one paragraph
Poppy is the nickname of the Personal Agent Protocol, an open
(Apache 2.0) standard led by Meta and Sierra. A company publishes /.well-known/poppy.json. It says where the
company's own agent is (a conversation endpoint), which APIs a personal agent may call (MCP or OpenAPI), whether to browse
the website, and how the shopper signs in. A personal agent such as Meta's Muse identifies itself with signed headers on
every request and asks the shopper before it shares anything or makes a change. Shopping itself (catalog, cart, checkout)
is delegated to UCP through an extension co-written with Shopify.
Three versions in ten days
| Version | Date | Status in the spec |
|---|---|---|
| 0.3 | 4 November 2026 | "Draft version 0.3. Any part of it can change." |
| 0.4 | 10 November 2026 | "The 1.0 release candidate", fixes may still break compatibility |
| 1.0 | announced 13 November, frozen 14 November | "Version 1.0. It's 0.4 with the fixes implementers needed." |
Before 1.0, two versions are compatible only if major and minor match. So 0.3, 0.4 and 1.0 are three different, mutually incompatible protocols on the wire. From 1.0 on, a newer minor version won't break older agents.
0.3 to 0.4: almost everything on the wire changed
- Versions: new
supported_versionsmap in the file, newPoppy-Versionheader on every request, newversion_unsupportederror. - Extensions: each one becomes a list (one entry per version), and the agent declares the ones it uses per
request in a
Poppy-Extensionsheader instead of in its client metadata. - Tasks:
feedback,excluded_tasksandskill_indexfold into onetasks.endpoint. - Conversations:
statusbecomesworking/handoff/closed, and every request carries anid. - Operations (the approval flow for changes such as orders) goes from v2 to v3, which is incompatible:
v3 adds approvals by the agent on the user's standing instruction, makes
termsrequired and lets a company decline an approval.
0.4 to 1.0: small, but still breaking
- Every entry in
agent.protocolsnow needs adescription. - Companies must accept PS384 signatures; client assertions live at most five minutes.
- Sign-in tightened: optional pushed authorization requests,
prompt=consentfor Direct Sign-In, the same DPoP key at refresh. - The other file fields, the headers and the Session token are unchanged.
Why the version number matters in practice
A company serves its preferred version at poppy.json and links the others from
supported_versions. The agent picks the newest one it supports. An agent that only speaks 0.4 and finds a
1.0-only file must not use it at all, and a 1.0-only agent can't use today's Shopify files. Publishing one version is
enough to be invisible to the agents that speak another.
Who publishes poppy.json today
We fetched /.well-known/poppy.json on 15 November 2026 from 112 domains of the companies listed as design
or live partners, from 12 large Shopify stores and from 11 Wix sites.
| Group | Serving a file | Version | What's in it |
|---|---|---|---|
| Design and live partners | 6 of ~105 companies | 0.3 (Rocket, DIRECTV), 0.4 (Mercury, Turkish Airlines, GoodRx, GoFundMe) | MCP APIs on the company's own domain; sign-in for some. No 1.0. |
| Shopify stores | 11 of 12 | 0.4, no supported_versions | One ucp
agent entry (no description), a checkout scope. No APIs, no sign-in server. |
| Wix sites | 9 of 11 | 1.0, plus 0.4 via supported_versions | An MCP server on the site's own domain (site search, business details, starting a purchase). |
Walmart, Target, Nordstrom, Gap, Wayfair, HelloFresh, United, Delta and Hyatt, all on the partner list, served no file. Neither did Synchrony, which Sierra named as live. A few partners sat behind bot walls we didn't try to get around. Every API endpoint we found was on the company's own domain; none was hosted by Sierra.
What Muse actually speaks
Muse publishes its client metadata at muse.ai/poppy/agent.json. It declares the operations extension at
version 2 inside the metadata itself, which is the 0.3 way of doing it (0.4 moved that to a header and removed the field).
On our server, Muse's requests carried Poppy-Version: 0.4, a valid signed Poppy-Session, and a user
agent of Muse-PersonalAgent/0.3. When we asked, Muse told us it supports 0.3 and 0.4. In practice:
- It reads 0.3 and 0.4 files.
- It ignored a file that offered only 1.0, and just browsed the site.
- It offers orders through operations v2, so a company that only accepts v3 won't get them.
Muse and Dot on the same sites
We asked Meta's Muse and Dot by OpenAI the same questions on five sites, each in a fresh chat, and asked each agent afterwards how it got the answer. On our own store we also read the server logs. OpenAI is a design partner too, on the agent side with Meta, Manus and Instinct.
| Site (what its poppy.json offers) | Muse | Dot by OpenAI |
|---|---|---|
| Rocket Mortgage (0.3, MCP API) | Used Rocket's MCP tool, after asking the shopper | Tried to read the file, its web tool couldn't open it; read the rates page |
| Wix store (1.0 + 0.4, MCP server) | Used the store's MCP tools: 6 of 10, 8 minutes, about 8 approval prompts | Read the product pages, about a minute |
| GoodRx (0.4, sign-in only) | Browsed (nothing else on offer); found a local price | Browsed; couldn't set the ZIP code, so no local price |
| Large Shopify store (0.4, UCP only) | Read the file, then browsed every colorway for four minutes | Used Shopify's in-page WebMCP tools (catalog search, product, variant), about two minutes |
| Our WooCommerce test store (1.0 + 0.4 + 0.3, MCP and a store agent) | MCP tools, a chat with the store agent, a test order | Read the product page; never requested the file (server logs) |
Both agents got the facts right on four of five sites; Dot couldn't set GoodRx's ZIP code. The difference is the route: Muse goes through Poppy when a site offers tools it can use. Dot never used Poppy, but it did use structured tools when they were in the page itself, which is how Shopify stores serve them.
Rocket Mortgage (design partner, 0.3)
Asked for today's sample 30-year rates, Muse read Rocket's file (version 0.3, an MCP API at
rocket.com/poppy/mcp), asked the shopper for permission and answered through it. The numbers matched Rocket's
public rates page, which is what that MCP tool says it serves. Dot gave the same numbers from the rates page; it tried
to open Rocket's poppy.json, but "the web tool couldn't access it".
A Wix store (1.0 with a 0.4 fallback, Wix's MCP server)
Asked which wheelchair wheel covers fit a 24-inch wheel and what a pair costs, Muse said up front it would go "through the store's own Poppy agent/API". Wix offered ten tools; Muse used six, from a site search and business details to a raw catalog query, and came back eight minutes later with all 225 designs at €169 a pair. Along the way it asked the shopper about eight times to "share information" with the store, some prompts describing a read-only question as a change to "your store data". Dot read the catalog and FAQ pages and had the same price in about a minute.
GoodRx (0.4, sign-in only)
GoodRx's file offers a sign-in server and nothing else, so Muse browsed: it got past a "press and hold" human check by starting a fresh browser, then compared pharmacy prices page by page and found the lowest local one. Dot also browsed, but the page stayed on another ZIP code and Dot said it couldn't verify a New York price.
A large Shopify store (0.4, UCP only)
Asked which men's sneakers were in stock in size 10, Muse opened the store's poppy.json, noted the UCP
entry, then clicked through every colorway in a browser for about four minutes. In its own words: "the answer came from
browsed pages, not from the store's own agent or its APIs. No Poppy/UCP call was made." On Shopify today, Poppy tells
Muse where UCP is, and Muse doesn't follow it yet. Dot took a different route to the same answer: in its browser
the store's pages expose Shopify's WebMCP tools, and Dot used them (search_catalog, browse_store,
get_product with size 10, show_variant) and finished in about two minutes. That isn't Poppy,
but it's the store's own structured data.
Our WooCommerce test store (1.0 + 0.4 + 0.3, MCP and a store agent)
We built a Poppy endpoint for WooCommerce stores and changed one thing at a time, watching the logs:
- 1.0 only: Muse browsed. No Poppy calls.
- 1.0 plus a 0.4 document, endpoints on our own API host: Muse tried, but its browser mistyped our host name and wandered off.
- MCP API on the store's own domain: Muse called it within minutes (
search_products, thenget_product), after asking the shopper "Allow Muse to share information with" the store. - Conversation endpoint on the store's domain: Muse opened a real Poppy conversation, showed it as a chat with the store, and quoted our store agent's shipping answer.
- Orders through operations v2: Muse called
create_order, showed the order as an approval card, sent the approval when the shopper said yes, and showed the store's payment link for the pending order. The store never touched card details.
Dot answered the same stock and shipping question correctly from the product page: "I didn't use the store's agent
or call any store APIs." Our logs showed no request for poppy.json during its run, while a test fetch of our
own showed up within seconds.
Is it usable?
Yes, if you build for the agent that exists rather than the spec that was just frozen. What worked for us:
- Serve 1.0 and 0.4 (and 0.3 if you can). Link the older documents from
supported_versions. - Lead with MCP tools. Muse reached for
apisbefore the conversation agent. Product search, product details, shipping and returns cover most questions. - Keep every endpoint on the store's own domain. Vendor hosts got mistyped or distrusted.
- Accept operations v2 for orders until Muse moves to v3.
- Expect caching: Muse kept an old tool list, so we version the MCP address whenever the tool set changes. It also lists tools without signing in, so list everything and check the session when a tool runs.
- Verify the Session header (a signed JWT with a short-lived delegated key) instead of trusting a user agent string.
What it means for your store
| Shopify | Shopify already publishes a 0.4 file pointing to UCP, and you can't edit it or add your own agent. Nothing to do, but don't expect Muse to use it yet; your product pages still carry the answer. Dot did use Shopify's in-page WebMCP tools, so complete variant data pays off there too. Complete product data matters more than any file. |
|---|---|
| Wix | Wix publishes 1.0 with a 0.4 fallback and an MCP server for your site, automatically. |
| WooCommerce and custom stores | Nothing is published for you. This is where a store can get ahead: serve the versions Muse reads, expose MCP tools on your domain, and you're one of the few stores Muse can talk to directly. |
Scelto serves Poppy for WooCommerce stores in our current test build, with the store's own MCP tools, a store agent and order approvals. If you run a WooCommerce store and want to try it, write to hello@getscelto.com. The free agent check shows what agents can read on any store today.
How we checked, and the limits
- Spec: the versioned pages, changelog and extension pages at personalagentprotocol.org, and the public repository, read on 15 November 2026.
- Adoption: one fetch per domain (plus www. and app. variants and a retry with browser headers), 15 November 2026. A company can serve Poppy on a subdomain we didn't try.
- Muse and Dot by OpenAI: the same five sites, 14 and 15 November 2026. Each agent's description of its own method is its own account; on our store we confirmed it in server logs, on the other sites we couldn't see the server side. In the Wix run we allowed each approval once and declined the ninth, a repeat; Muse then answered from what it had.
- Muse changes often. We'll update this page when its version support changes.
Questions
What is Poppy?
Poppy is the short name of the Personal Agent Protocol, an open standard from Meta and Sierra for how personal AI agents such as Meta's Muse talk to businesses: a file at /.well-known/poppy.json lists the company's agent, APIs and sign-in.
Which Poppy version should a store publish?
As of 15 November 2026, publish 1.0 and link a 0.4 document from supported_versions. Muse sends Poppy-Version 0.4 and ignored a file that offered only 1.0 in our test.
Does Muse use Poppy on Shopify stores?
Shopify publishes a Poppy 0.4 file for every store that points to its UCP endpoint. In our test on a large Shopify store, Muse read the file and then browsed the product pages instead; it made no UCP call.
Does Dot by OpenAI use Poppy?
Not in our tests on 15 November 2026. OpenAI is a design partner, but Dot read pages on Rocket Mortgage, a Wix store, GoodRx and our WooCommerce test store, and never requested poppy.json on ours. On a Shopify store it used Shopify's in-page WebMCP tools instead.
Can I edit poppy.json on Shopify?
Not as of November 2026. Shopify generates the file, and neither merchants nor apps can add their own agent or APIs to it.
See how AI agents read your store
The free agent check works on any store platform and needs no sign-up.
Check your store