Personal Agent Protocol (PAP) Explained: Sierra and Meta's OAuth-Based Standard for Agents and Businesses vs MCP, AP2 and x402
PAP (Poppy) Draft 0.1 is published: poppy.json discovery, DPoP-bound OAuth sessions, direct, device and mediated sign-in, MCP and OpenAPI routes, and how it differs from AP2 and x402.
The Personal Agent Protocol (PAP, also called Poppy) is an open standard for how a person's AI agent works with a business. Sierra announced it on October 6, 2026, with Meta co-developing and Genesys, Instinct, Rocket, Shopify, Stripe and Walmart named as partners. One correction up front, because the acronym invites it: PAP is not a payment protocol.
Updated October 9, 2026. When we first published this post, the specification was unreleased. The Draft 0.1 specification is now live at personalagentprotocol.org (last updated October 9, 2026), so this post now describes what the draft actually says. The project site calls every part of it changeable, "including the parts the specification already covers."
Sourcing: the announcement details (Sierra, Meta, partners) come from press coverage by The Next Web and Unite.AI. Everything about the protocol itself comes from the draft specification and guides. The site itself names no organization or governing body.
The problem PAP targets
The overview describes today's choice as binary: agents either pass as the user by signing in with the user's password, or companies block them. PAP aims for a middle ground where "companies and users can choose how agents work with them." Its five stated principles are building on existing standards, controls rather than policy, security and privacy, progressive adoption, and extensibility.
What the draft defines
| Piece | What Draft 0.1 says |
|---|---|
| Discovery | A company publishes https://{domain}/.well-known/poppy.json with protocol_version, organization, auth, agent, apis, web and extensions |
| Authorization server | RFC 8414 metadata including issuer, token_endpoint, revocation_endpoint, and a poppy_domains list |
| Session token | Short-lived access token for one session, bound to a DPoP key |
| Account token | Long-lived refresh token for future signed-in sessions |
| Starting a session | The agent sends a signed JWT assertion (RFC 7523 JWT Bearer grant) naming its client_id, a per-company user ID, and the company's token endpoint |
| Sign-in methods | Direct (OAuth authorization code with PKCE), Device (device authorization grant), Mediated (agent submits credentials) |
| Scopes | poppy:read, poppy:write, plus custom scopes declared in auth.custom_scopes |
| Agent identity | The agent publishes a client metadata document at its client_id URL with jwks_uri and private_key_jwt authentication |
| APIs | OpenAPI 3.0/3.1 (DPoP tokens) or MCP servers (bearer tokens, MCP 2025-06-18 or later) |
| Web | The agent joins its browser to a session with a signed assertion; the company sets a Secure, HttpOnly, SameSite=Lax cookie |
| Conversations | HTTPS endpoints for messages, event streaming, handoff to a person, and close |
The session ladder, concretely
The announcement describes agents starting as guests and a customer then choosing read-only or write access. In the draft, the token response carries a signed_in flag recording whether a user has signed in, a customer can sign in by one of the three methods, and the company decides which scopes to grant. User IDs must be stable per company but opaque and unique across companies, so two companies cannot match the same person. The site's homepage also describes five permission levels, from blocking agents entirely to full autonomy, with read-only access and change approval in between.
Why DPoP matters here
Session tokens are bound to a key the agent holds. Each request carries a proof signed with that key (method, URL, a unique ID and a hash of the token), so a copied token is useless without the private key. The one exception is MCP: because MCP clients expect bearer tokens, the draft allows bearer session tokens on MCP servers only, and flags them as less secure. If you want to check whether an authorization server advertises DPoP, the OIDC Inspector shows the DPoP algorithms in its published metadata.
The operations extension
Draft 0.1 defines one extension, operations (version 1): "The personal agent confirms an action before the company performs it." Approval covers the exact terms, the action has one record across channels, and a retry never performs it twice. Extensions are optional; a party that does not implement one ignores it and works under the core spec.
What PAP does not cover (yet)
The open topics page lists three gaps:
- Payments. "How a personal agent makes payments in a way that fits the protocol, including how that works with existing payment protocols." Nothing is specified yet.
- Push notifications. How a company reaches the agent when no request is open, for example when an order ships.
- Attachments. Receipts, return labels, images or forms inside company messages.
Other things to keep in view:
- Draft status. Draft 0.1 can still change.
- Governance. The specification, documentation and examples are under the Apache License 2.0, which excludes the names "Personal Agent Protocol" and "Poppy" and the logo. The site names no steward or standards body.
- Production use. We found no reported deployments.
PAP vs MCP vs AP2 vs x402
These answer different questions, and the mistake to avoid is treating them as competitors.
| Protocol | Question it answers | Layer |
|---|---|---|
| MCP | What tools and data can an agent call? | Tool and API interface |
| PAP | Who is the agent acting for, and what may it do at this business? | Session and authorization |
| AP2 | Did the user authorize this specific purchase? | Payment authorization (Intent, Cart and Payment mandates) |
| x402 | How does an agent pay an HTTP endpoint? | Payment settlement over HTTP 402 |
A plausible stack is PAP for the session and permissions, MCP or OpenAPI for the calls, and a payment protocol for money once PAP's payment topic is specified. That last link is our reading, not something the spec says. For the payment layers, see what x402 is and our x402 pre-signing checklist.
What to do now
- Audit your authorization server metadata. PAP requires RFC 8414 metadata with specific endpoints and a
poppy_domainslist. Run your issuer through the OIDC Inspector to see what it publishes today, and use the Poppy Inspector to checkpoppy_domainsand the endpoints your sign-in methods need. - Read the company guides. They recommend session tokens that last hours, account tokens of about 90 days, and PKCE with
S256on consent pages. - Decide your scopes and your guest tier. Which read-only actions are safe without sign-in, and which custom scopes would you define?
- Inventory your agent-facing surface. Know which endpoints already expose MCP or OpenAPI. The Agent Protocol Inspector shows what an agent can discover about an MCP server, A2A card or ARD catalog, and its Poppy check reads and lints a
poppy.jsonand cross-checks its OAuth issuer. It does not call the endpoints apoppy.jsonlists. - Watch the spec. Payments, notifications and attachments are open, and the draft may change.
If your OAuth setup is the weak link, our guide to MCP 401 and OAuth discovery covers the same machinery.
Sources
Follow Trango Compute on LinkedIn
We post updates on new tools, context engineering patterns, and LLM cost research.