Review · November 16, 2026
Personal Agent Protocol (Poppy) 1.0 Summary
The problem it's solving
A personal agent acts for one user at many companies. Today the company either lets the agent pretend to be the user, typing the user's password into its website, or it blocks the agent.
Poppy defines the middle path. The agent always identifies itself. The company recognizes a returning user by a pseudonymous ID before they sign in. When a task needs the account, the user grants scoped access through OAuth.
Version 1.0 was published on November 14, 2026, five weeks after 0.1. It is frozen, and 1.x versions promise not to break it.
How does it fit in with OAuth?
This is not an IETF document. It is an independent spec under Apache 2.0, and it never went through a standards body.
The 1.0 design is a much cleaner OAuth profile than 0.1. Almost everything 0.1 added to the authorization server is gone:
- The JWT bearer "session grant" is gone. Sessions are now a
Poppy-SessionJWT header that the agent sends to resources. The authorization server never sees them. session_id,signed_inandpoppy_domainsare gone. The domain-to-issuer check usesprotected_resourcesfrom Protected Resource Metadata.- Scopes are the company's own. The protocol defines one,
poppy:browser-sign-in. - The password-style sign-in moved out of the core into an extension with a retirement date.
What the authorization server has to support is now ordinary: authorization code with PKCE, optionally the device grant and PAR, refresh, revocation, Client ID Metadata Documents, and private_key_jwt with ES256 and PS384.
Everything else sits above OAuth: discovery, sessions, web browsing, a conversation protocol, task coverage, and six extensions.
How it works
Discovery. The company publishes /.well-known/poppy.json. It lists the company's APIs (OpenAPI or MCP), its own agent's conversation endpoint, a browser-session endpoint, and scope descriptions. The agent finds the issuer through the domain's protected resource metadata, and the issuer must list the domain back.
Agent identity. The agent's client_id is a CIMD URL. Redirect URIs and jwks_uri must be on the same domain. Every user of an agent shares that one client_id.
Sessions. Every request carries a Session JWT naming the agent, a pairwise User ID the agent made up, and a session ID. It is signed with a short-lived delegated key, and a key in the agent's jwks_uri vouches for that key in an embedded delegation JWT. The Session JWT also covers the method, URL and access token hash, like a DPoP proof.
Signing in. Direct Sign-In is the authorization code flow and is preferred. Device Sign-In is the device grant. Both return a refresh token.
Web browsing. The agent's browser loads a company page that hands it a one-time code. The agent authorizes that code from its own backend with the Session JWT and access token. The company then sets its own cookie. The access token never reaches the browser.
Approvals. With the Operations extension, a company refuses a consequential request and returns exact terms. The agent sends the same request again with an approval, marked approver: user or approver: agent.
What 1.0 gets right
The OAuth choices are stricter than RFC 9700 in several places:
- Exact redirect URI matching, and
issrequired on every authorization response. Agents talk to thousands of authorization servers, so mix-up defense matters here. - Client assertions are addressed to the issuer, following rfc7523bis, and last at most 5 minutes.
scopeis required, and the company MUST NOT grant more than was requested.- The consent page MUST list every scope being granted, including ones granted before.
- A refresh must use the same DPoP key the refresh token was issued with, and the company MAY bind the refresh token to it.
The last two are new in 1.0. Both exist because every user of an agent shares one client_id, which is a problem most OAuth deployments never had to think about.
Significant concerns
The client is confidential in name only. The spec says an agent SHOULD sign client assertions blindly, so its own servers can't see which companies a user works with. It then admits that "any of the Personal Agent's Users can authenticate as the Personal Agent."
That makes client authentication worth very little. RFC 9700 requires refresh tokens for public clients to be sender-constrained or rotated. Poppy makes DPoP binding of refresh tokens a MAY and leaves rotation to the Guides.
Revoking the client_id is the only enforcement, and blind signing undercuts it. Several rules can't be verified: the user must sign in themselves, and approver: user must be true. The remedy each time is that the company revokes the agent's client_id. An operator that signs blindly can't see or stop what happens under its own client_id.
Consent phishing is an accepted risk. Anyone can start a sign-in and send the link to someone else. Section 10.8 says so directly. Classic OAuth clients stop this by tying state to the browser session. Poppy only says the agent SHOULD receive the redirect on the device that started the flow.
One scope can swallow the rest. If a website can't enforce scopes, poppy:browser-sign-in means everything the account can do there. The company must say so on the consent page. A user who approves "see your orders" plus browser sign-in has still approved full access.
Approvals are the agent's word. The company can't verify either kind of approval. The path where the user confirms on the company's own page is explicitly not defined. Signed approvals are on the roadmap.
Mediated Sign-In is still the password grant. The agent sends the user's password, and relays one-time codes, to the token endpoint. It can't use passkeys and has no consent page. It retires on March 17, 2027, but "the community will look for ways to encourage" companies to move after that. That is not a removal.
Sender-constraining is optional twice. DPoP is the company's choice. Verifying the Session JWT is a SHOULD. A company that skips both accepts plain bearer tokens.
Significant interoperability concerns
The OAuth parts are tightly specified. 1.0 fixed the biggest gap from 0.4 by making PS384 mandatory, so agents that sign blindly are no longer refused by conforming companies.
Three areas will cause trouble:
- Four proofs on one request. A signed-in request can carry an access token, a DPoP proof, a Session JWT with an embedded delegation, and a Web Bot Auth signature. Each has its own key and its own verification rules.
- Cross-device rules can't be tested. Both sign-in types "MUST follow" RFC 10027 §2, which asks for a risk assessment and "appropriate mitigations".
- Nothing is registered.
Poppy-Session, thepoppy_dlgJOSE header, the JWTtypvalues and/.well-known/poppy.jsonhave no IANA entries. The spec has no IANA section.
Mechanisms that could be removed
- Mediated Sign-In. It has a date. Remove it on that date.
- The conversation protocol. It is a full chat protocol with events, streaming, handoff and state.
agent.protocolsalready allows other types, so this could be a profile of an existing agent protocol. ath,htmandhtuin the Session JWT, for companies that require DPoP. They duplicate the DPoP proof.
Writing quality
Very good. Plain English, consistent terms, correct BCP 14 keywords, and an example for nearly every message. Section 4.8 lists everything Poppy adds to OAuth in one place. More profiles should do that.
The security considerations are candid about what the protocol accepts as risk.
The weakness is the split between the spec and the Guides. S256-only PKCE, refresh token rotation, and most of what the consent page must show are in the Guides. The Guides open by saying that following them "is not required to comply with the protocol."
It is also long. The spec, guides and extensions run to several thousand lines, and it went from 0.1 to a frozen 1.0 in five weeks.
Background of the authors
The spec names no authors, so there are no Datatracker profiles to check.
The site lists 103 design partners. Meta, OpenAI, Manus and Instinct are the personal agents. The rest are companies and technology providers, including Okta, Ping Identity, Cloudflare, Salesforce, Stripe, Shopify, Visa, Mastercard, PayPal, Walmart and Delta.
With that list, this ships whether or not it ever reaches a standards body. Governance is still not written down beyond the license and the contribution terms.
Suggested feedback
- Require DPoP-bound or rotated refresh tokens whenever user devices can obtain client assertions.
- Require the agent to match the browser at
redirect_urito the user who started the sign-in, whenever the agent has its own login. - Move the consent page requirements from the Guides into the spec: the
client_iddomain, the account, and the "only if you just asked" warning. - Make DPoP or Session JWT verification mandatory for signed-in requests.
- Define the "approve at the company" step. Rich Authorization Requests already does this at the authorization endpoint.
- Bring the Session JWT and delegated key design to the OAuth or WIMSE working groups, and register the headers and the well-known URI.
Further reading: