Updated 16 October 2026. The specification now has a public GitHub repository (since 14 October), and feedback goes through its issues. The first round of fixes landed on 15 and 16 October, still as draft 0.1. We've updated the governance section and sources.
Updated 13 October 2026. The protocol site added a Path to 1.0 page, a feedback route and its design-partner list on 12 October; the specification text didn't change. We've updated the governance section and links.
Updated 11 October 2026. Rewritten now that the draft 0.1 specification is published (9 October). The first version of this post, from 8 October, covered the announcement only.
The Personal Agent Protocol now has a specification. Meta and Sierra announced it on 6 October. Three days later they published draft 0.1 at personalagentprotocol.org and named 35 design partners, including OpenAI, Visa, Mastercard, PayPal, Cloudflare and Target.
The spec answers most of the questions the announcement left open. A business publishes a discovery file at /.well-known/poppy.json. Sign-in is OAuth. A customer grants poppy:read, poppy:write or both. One session then covers the business's APIs, its website and a conversation with its own agent. This post covers what the draft defines and who is involved. It also covers where PAP sits next to the Universal Commerce Protocol, keeping to what has been published.
What is the Personal Agent Protocol?
The Personal Agent Protocol (PAP, nicknamed "Poppy") is an open protocol for how a person's AI agent works with a company on their behalf. It covers how the agent finds the company, who it is acting for, what it is allowed to do, and how it gets the work done through APIs, web pages or the company's own agent. The introduction puts it this way: "A company uses it to declare how personal agents can interact with it. A personal agent uses it to sign in safely and get access to a person's account, limited to what that person and the company agree to."
The specification and docs are licensed under Apache 2.0. The names "Personal Agent Protocol" and "Poppy" and the logo are not covered by that licence.
The problem it targets
The draft describes today's choice as "all or nothing". Either the agent passes as the user, "signing in with their password and clicking through pages built for people", or the company blocks it. Neither side can set limits in between. Companies spend effort detecting agents, and agents spend effort avoiding detection.
PAP's answer is a range of options between those extremes. The spec's own example is an airline. Any agent can search flights. A signed-in agent can see the user's trips. Buying a ticket needs the user's approval. The user, for their part, might let an agent see trips but not change them.
How draft 0.1 works
The specification has five parts, plus extensions.
1. Discovery: /.well-known/poppy.json
A company publishes a JSON document at https://{domain}/.well-known/poppy.json. Here it is, cut down from the spec's example:
{
"protocol_version": "0.1",
"organization": { "name": "Example Company", "domain": "example.com" },
"auth": {
"issuer": "https://auth.example.com",
"direct": { "scopes": ["poppy:read", "poppy:write"] }
},
"agent": {
"protocols": [
{ "type": "poppy", "endpoint": "https://api.example.com/poppy/conversations" }
]
},
"web": { "browser_session_endpoint": "https://example.com/poppy/browser-session" },
"apis": [
{ "type": "openapi", "url": "https://api.example.com/openapi.json", "description": "Orders, returns, and exchanges" },
{ "type": "mcp", "url": "https://mcp.example.com/mcp", "description": "Product search and sizing" }
]
}
The rules that matter:
organization.domainmust match the domain the agent asked for.- A company offers at least one of three routes: a company agent, APIs, or its website.
- The company's OAuth server must publish standard metadata (RFC 8414) with a new
poppy_domainslist naming every domain allowed to use that issuer. This stops one site claiming another company's sign-in server.
2. Sessions and sign-in
Everything happens in a session: one user, at one company, through one personal agent.
- Identifying the agent. The agent identifies itself with a
client_idthat is a URL to its own metadata document and public keys. This is the same OAuth Client ID Metadata Document draft that MCP uses. - Before sign-in. The company sees only an ID for the user that is stable but different at every company, and that "MUST NOT contain or be derived from personal information".
- Token security. Session tokens are short-lived and bound to a key the agent holds (DPoP, RFC 9449), so a copied token is useless on its own.
When a task needs the user's account, the user signs in in one of three ways:
| Type | Who signs in | How |
|---|---|---|
| Direct | The user | On the company's own page, via OAuth authorization code with PKCE |
| Device | The user | On any device, with a link and code the agent shows them (OAuth device grant) |
| Mediated | The agent, for the user | The agent sends credentials the user gave it to the company's sign-in endpoint (not OAuth) |
Access is expressed as scopes. poppy:read means view the account, such as orders and bookings. poppy:write means make changes, such as an exchange. Companies can add their own scopes, or offer only narrower ones. Each sign-in type lists the scopes it can grant.
Sign-in gives the agent an account token: an OAuth refresh token, tied to that agent's keys. It lasts until the user signs out, disconnects the agent from their account settings, or the company revokes it.
3. Web browsing
If a company lists a browser_session_endpoint, the agent's own browser joins the session by posting a signed assertion. The company then sets its own cookie and applies the session's scopes on every page. Browsing, API calls and conversation all show up as one visit.
4. APIs
A company can list OpenAPI descriptions and remote MCP servers. PAP leaves each API's schema to that API. The protocol only requires that every listed API accepts its session tokens. MCP servers get audience-bound Bearer tokens, because MCP can't use DPoP. MCP clients that don't implement PAP can still sign in through MCP's own flow.
5. Conversations
PAP defines its own JSON-over-HTTPS conversation protocol with the company's agent. It supports sending messages, reading events (with long polling or server-sent events), handing off to a person, and closing. Every message says whether a person or an AI wrote it, on both sides. A user can also step into a Direct Conversation and talk to the company themselves through their agent's interface.
Extensions: operations
The first extension, operations, lets a company propose an action (an exchange, a booking change, a cancellation) with exact terms. The agent confirms that exact revision, with the user's approval where needed. The action keeps one ID across API, MCP, conversation and web, and a retry never performs it twice.
What the draft leaves out
The open topics list three areas the draft leaves out on purpose:
- Payments, "including how that works with existing payment protocols";
- push notifications;
- attachments such as receipts and return labels.
The spec says any part "can change before a stable version, including in ways that aren't backward compatible".
Who's involved
Launch partners (Sierra, 9 October): Meta, Sierra, Genesys, Instinct, NiCE, Rocket, Shopify, Stripe and Walmart. The 6 October announcement quoted four of them:
- Shopify (Mani Fazeli): the protocol "helps establish how agents and merchants can work together".
- Stripe (Kevin Miller): "A customer relationship doesn't start or end at checkout."
- Genesys (Tony Bates): "Brands need a trusted way to know who an AI agent represents, what it's authorized to do, its intent, and how to work with it securely."
- Rocket Companies (Shawn Malhotra), on a Muse agent working across its home-buying platform.
Design partners (35, added 9 October; since 12 October the site lists them together with the launch partners):
- Payments and finance: Adyen, Atomic, Bank of America, BBVA, Chime, Mastercard, PayPal, Plaid, Synchrony, Venmo, Visa and Wells Fargo.
- Retail: Gap, Nordstrom and Target.
- Travel: Hertz and United Airlines.
- Insurance and health: Cigna, GEICO, Insurify and Liberty Mutual.
- Media and telecom: Comcast, DIRECTV, FOX and SiriusXM.
- AI and software: OpenAI, Cloudflare, ElevenLabs, Klaviyo, Notion, Okta, OneSignal, Zapier, Zendesk and 1Password.
OpenAI wasn't involved at the 6 October launch and joined as a design partner three days later. Anthropic, Google, Amazon and Apple aren't named.
Governance is filling in. Since 14 October the specification and docs site have a public GitHub repository, under Apache 2.0, with issue tracking and revision history, and the site links to it. The site still doesn't name its authors; the commits so far come from Sierra. A Path to 1.0 page, added on 12 October, says the authors are working through implementations with several design partners and that revisions before 1.0 can break compatibility. Feedback now goes through GitHub issues, written by hand, from the site's exported inline comments or by an agent using a downloadable kit of the docs, under published contribution terms. Changes are already landing: the first spec fixes, from sign-in errors to a new security considerations section, were merged on 15 and 16 October, still as draft 0.1. Sierra's post asks companies to "read the draft and help shape this new protocol".
Builders started straight away. Within two days of the draft:
- Cloudflare's Brendan Irvine-Broque published a two-Worker demo of a personal agent and a receiving store.
- Datalayer released a Python and TypeScript SDK.
- poppyseed appeared as a Next.js library for sites.
How it relates to UCP
The PAP spec doesn't mention UCP, and UCP doesn't mention PAP. Sierra has described the two as complementary. With the draft published, the comparison can now go beyond the announcement.
| UCP | PAP draft 0.1 | |
|---|---|---|
| Discovery file | /.well-known/ucp |
/.well-known/poppy.json |
| What it defines | Commerce operations: catalog, cart, checkout, orders, payment handlers, verticals | Identity, sessions, permissions and conversation, for any kind of company |
| Signing in | Identity linking (OAuth 2.0), with capability-scoped tokens | OAuth core, plus device and mediated sign-in; scopes poppy:read / poppy:write |
| Agent identity | Web Bot Auth interop (signed requests) | client_id metadata URL with published keys; DPoP-bound tokens |
| Transports | REST, MCP, A2A, embedded | OpenAPI, MCP, its own conversation protocol, web via session cookie |
| Payments | Payment handlers | Open topic |
Where they can meet. PAP's apis list takes any OpenAPI or MCP endpoint and leaves the schema to that API. A store that already runs a UCP MCP server could list it there. The commerce operations would then be UCP's, and the session around them would be PAP's. Nothing in either spec says this is the intended pattern, but nothing in draft 0.1 rules it out.
Where they overlap. Both have discovery files. Both use OAuth sign-in. Both answer guest questions without a sign-in: PAP through a signed-out session and conversation, UCP through the ask capability. A store running both would publish two discovery documents and might run two OAuth integrations.
The same companies are in the room. Shopify and Stripe are PAP launch partners and active in UCP. Meta staff co-chair UCP's Bulk Product Discovery working group.
What it means for stores and agent builders
-
Stores: nothing is required yet. It's a 0.1 draft with no stable version or versioning policy. If you already run OAuth for customer accounts and an MCP server, most of a
poppy.jsonalready exists in pieces: an issuer, scopes and an API URL. We found nopoppy.jsonon the launch partners' main domains on 11 October, which is expected two days after a draft. To see how yours reads to an agent, PAP Checker validates a domain'spoppy.json, its sign-in server and the APIs it lists against draft 0.1, and keeps a list of sites publishing one. -
Agent builders: a client needs:
- a published
client_idmetadata document and keys; - a per-user, per-company ID;
- DPoP;
- the three sign-in flows;
- the conversation protocol.
The guides add advice on keeping tokens and private keys out of model context.
- a published
-
Commerce: checkout and payment stay outside PAP for now. If your agent buys things, it still needs a commerce protocol such as UCP for the cart and the order.
What we'll watch
- Reference implementation and workshops: Sierra says both arrive "over the next month".
- A conformance suite or JSON Schema: neither has been published yet.
- The payments extension: how it relates to UCP payment handlers and existing card network work.
- First live
poppy.jsonfiles from launch partners (tracked at papchecker.com/sites). - Whether any UCP business lists its UCP endpoint in
poppy.json. - Governance: named maintainers, and how the spec moves from draft 0.1 to a numbered revision. A path to 1.0 arrived on 12 October and a public repository on 14 October.
FAQ
What is the Personal Agent Protocol? An open protocol, nicknamed Poppy, for how a person's AI agent works with a company on their behalf: discovery, sign-in, what the agent may do, and how it works through APIs, web pages or the company's own agent. Meta and Sierra lead it, and draft 0.1 was published on 9 October 2026.
What is poppy.json?
The discovery document a company publishes at /.well-known/poppy.json. It names the organisation, its OAuth issuer and sign-in types, and the APIs, website session endpoint and company agent it offers.
Is the Personal Agent Protocol specification available? Yes. Draft 0.1 is at personalagentprotocol.org under Apache 2.0. It's a draft, and any part can change before a stable version.
Who backs the Personal Agent Protocol? Meta and Sierra, with launch partners Genesys, Instinct, NiCE, Rocket, Shopify, Stripe and Walmart. There are also 35 design partners, including OpenAI, Visa, Mastercard, PayPal, Cloudflare, Target and United Airlines.
Does the Personal Agent Protocol use OAuth? Yes. Sessions and sign-in are OAuth 2.0, with DPoP-bound tokens, PKCE and the device grant. The exception is "mediated" sign-in, where the agent submits credentials the user gave it.
Is the Personal Agent Protocol a competitor to UCP? They define different layers. UCP defines commerce operations such as catalog, cart and checkout. PAP defines sessions, sign-in and permissions, and leaves each API's schema to that API. They overlap on discovery and sign-in, and neither spec references the other.
Does the Personal Agent Protocol handle payments? Not in draft 0.1. Payments is listed as an open topic.
About UCP Checker
UCP Checker is the independent validation and observability layer for the Universal Commerce Protocol. We crawl, validate and grade every public UCP manifest we can find, run the merchant directory, the UCP Score, live adoption stats, the vertical guides and the authority-bound vendor map, and track the spec as it evolves so you don't have to — measured from two vantages (what a business declares, and what actually happens when an agent transacts), the same way for everyone, without picking winners.
Sources
- Personal Agent Protocol specification, draft 0.1, with changes merged through 16 October 2026
- personal-agent-protocol/personal-agent-protocol, GitHub, public since 14 October 2026
- Personal Agent Protocol introduction, operations extension, guides, path to 1.0 and contribution terms
- Sharing a draft of Personal Agent Protocol, Bret Taylor and Clay Bavor, Sierra, 9 October 2026
- Introducing Personal Agent Protocol, Bret Taylor and Clay Bavor, Sierra, 6 October 2026
- Meta, Walmart, Instinct, Shopify, Sierra, Stripe, and others publish the Personal Agent Protocol, Kate Rooney, CNBC (via Techmeme), 6 October 2026
- poppy-demo, personal-agent-protocol SDK, poppyseed, GitHub
- Identity Linking capability (draft docs) and Ask capability (draft docs), ucp.dev
Related coverage: UCP's Ask Is Merged · Agent Identity in UCP: Web Bot Auth Interop · UCP's Bulk Product Discovery Working Group · Google's UCP Integration Hub
