By · Founder, Stacktree · Last updated
Personal Agent Protocol · Version 1.0

The Personal Agent Protocol, from a company that runs it.

Version 1.0 of the Personal Agent Protocol, Poppy for short, was published on 14 November 2026. Most explainers still describe the October drafts, and a lot changed between them. Stacktree has served every version since the first draft, so this is the protocol as it works today: what it is, each step on the wire, what 1.0 changed, and what it cost us to keep up.

The Poppy Simulator. Free, no sign-up.

What is the Personal Agent Protocol (Poppy)?

The Personal Agent Protocol is an open standard for how a person's own AI agent works with a company for them. The company publishes a poppy.json file saying what it offers. The agent identifies itself, starts a Session, has the person sign in on the company's own page when a task needs their account, and then uses the company's APIs, website or support agent. Sierra announced it with Meta in October 2026, and Version 1.0 was published on 14 November 2026.

What problem it solves

People are starting to hand errands to an AI agent of their own: return this jacket, move my booking, publish these notes somewhere my client can read them. To do that the agent has to deal with a company, and until now it had two bad options. It could drive the website like a person, which breaks on every login wall and looks like a bot. Or the company could build a one-off integration for each agent, which nobody has time for.

The Personal Agent Protocol gives both sides one agreed way to meet. The company decides what agents may do and says so in one file. The person decides what their agent may do for them, on the company's own consent page. The agent proves which agent it is on every request. Nobody shares a password, and the company can tell a customer's agent from a scraper.

How Version 1.0 works, step by step

Section numbers are from the 1.0 spec.

  1. Discovery (section 3). The agent fetches https://yourdomain/.well-known/poppy.json. The file names the company, its domain, and at least one of three channels: APIs, the website, or a Company Agent the personal agent can talk to. The agent checks that organization.domain is the host that served the file, so a copy on another domain cannot speak for you. We go through every field in a real file on its own page.
  2. Agent identity (4.1). An agent's client_id is the HTTPS address of a metadata document it publishes, with its name and its public keys. There is no registration step and no shared secret. The company fetches the document and knows who is asking.
  3. A Session, with nobody signed in (section 5). No request starts a Session. The agent signs a short-lived token for each request and sends it in a Poppy-Session header, with the version it speaks in Poppy-Version. A new session id is a new Session, signed out. Whatever the company allows signed out works from here. On Stacktree, that already covers publishing a page to a private link.
  4. Sign-in, when a task needs the account (4.2, 4.3). This is OAuth. With Direct Sign-In the person is sent to the company's own consent page, which must list every scope it will grant. With Device Sign-In, for an agent on glasses or a phone call, the person opens a link on any device and confirms a short code. The agent gets a refresh token and a short-lived access token. It never sees a password.
  5. Doing the task (sections 6 to 8). The agent calls an API the file listed, such as an OpenAPI description or an MCP server, with the Session header and, if signed in, the access token. Or its browser joins the Session and uses the website signed in. Or it opens a conversation with the company's own agent.
  6. When the company cannot help (section 9). A company can list tasks it chose not to offer, so the agent stops trying. And the agent can send feedback when a person asked for something the company has no way to do. That report is the cheapest product research a company will ever get.

From first draft to 1.0 in five weeks

The dates are from the protocol's own changelog. Each version is frozen at its own address, so an agent built for one can keep reading it.

  • Draft 0.1, 9 October 2026. The first published draft, three days after the announcement.
  • Draft 0.2, 24 October. Settled which domain a poppy.json speaks for, one OAuth issuer per company, and added feedback.
  • Version 0.3, 4 November. The big rewrite. Sessions were separated from OAuth and moved into a signed header. Companies define their own scope names. A company can serve agents signed out with no OAuth server at all.
  • Version 0.4, 10 November. The 1.0 release candidate. A company can serve several versions at once, and a browser joins a signed-in Session through its own scope.
  • Version 1.0, 14 November. 0.4 with the fixes implementers needed.

What changed in 1.0

The changelog lists seven major changes. The ones a company or an agent builder will feel:

  • Minor versions cannot break you. From 1.0, each side sends only what the other side's version defines. A 1.1 agent can use a 1.0 company, and the reverse. This is the change that makes it reasonable to build now.
  • The consent page lists every scope, every time. Every user of one personal agent shares a client_id, so a company that skipped the page for scopes granted earlier could hand one person's approval to a sign-in someone else started. Direct Sign-In now sends prompt=consent.
  • PS384 is required. Companies must accept agent signatures made with PS384 as well as ES256.
  • A company may require Pushed Authorization Requests. It is optional, and it lets banks and others on FAPI 2.0 take part.
  • Refresh uses the same key. An agent that sends DPoP proofs refreshes with the key it signed in with, and a company may bind refresh tokens to it.
  • Each conversation endpoint says what it helps with. An entry under agent.protocols needs a description, so a personal agent can choose between a support agent, a shopping agent and the APIs.

What the October explainers get wrong now

If you read about the protocol in October, four things you learned are no longer true:

  • "The agent requests a Session Token." Not since 0.3. There is no Session request and no Session Token. The agent signs a header for each request.
  • "Scopes are poppy:read and poppy:write." Companies name their own scopes now. Ours are sites:read and sites:write, the same ones our other connectors use.
  • "Tokens are bound to the agent's key with DPoP." DPoP is optional in 1.0. Plain Bearer access tokens are allowed, because the Session header already proves which agent is asking.
  • "poppy.json lists poppy_domains and the sign-in types." Both moved out. The link between a domain and its sign-in server is checked through standard OAuth metadata, and sign-in types are read from the authorization server.

Personal Agent Protocol and MCP

They answer different questions, and a company will often run both. MCP is how an agent calls tools on a server. The Personal Agent Protocol is everything that has to be true before that call is safe for a stranger's agent: which agent this is, which person it acts for, what they approved, and where your tools are. Our poppy.json lists our MCP server as one of two APIs. An agent that found us through the protocol lists our MCP tools signed out, with nothing but its Session header. The full comparison is on Personal Agent Protocol vs MCP.

What it took to run it

Stacktree hosts HTML pages at private links, and an agent publishing a page for its person is the whole job. We went live on Draft 0.1 on 11 October 2026, two days after it was published, and have followed every version since.

  • The first version was the expensive one. Agent identity, signature checks and a consent page are real work. Budget days, not hours.
  • 0.3 was a rewrite. We built it behind a flag while 0.2 kept serving, and switched the day after it was published.
  • 1.0 took two code changes. Accepting any 1.x in Poppy-Version, and one token request we had refused by mistake. Everything else in the changelog we already did or do not offer.
  • Old agents kept working. We still serve 0.4 and 0.3 beside 1.0, linked from supported_versions.
  • You need an agent to test against. Our own tests were written from our own reading of the spec. We found our real mistakes only after we built an agent from the spec text alone and pointed it at production.

We wrote up each move for the spec's maintainers: the reports for 0.1, 0.3 and 1.0 are public. If you are about to build it, the order we would do it in again is its own page.

Try it

  • Read a live file. stacktr.ee/.well-known/poppy.json is a 1.0 document serving real agents. The field guide explains each line.
  • Watch an agent. Open the Poppy Simulator, and it discovers Stacktree, starts a Session and publishes a page while you read every request.
  • Test your own. Choose Test a company in the simulator and enter your domain. It checks your discovery, a signed-out Session, your APIs and the requests you must refuse.
  • See who else runs it. Who supports Poppy lists the companies whose files we could read, with the version each serves.
  • Building an agent? Stacktree for agents lists what ours accepts, and llms.txt has the same in a form an agent can read.
FAQ

Frequent questions

What is the Personal Agent Protocol? +
An open standard for how a person's own AI agent works with a company on their behalf. The company publishes a file called poppy.json that says what it offers. The agent identifies itself, starts a Session, asks the person to sign in on the company's own page when a task needs their account, and then uses the company's APIs, website or support agent.
Is Poppy the same thing as the Personal Agent Protocol? +
Yes. Poppy is the protocol's short name, and the spec uses both. You will also see PAP. The discovery file is poppy.json, the request headers are Poppy-Session and Poppy-Version, and the site is personalagentprotocol.org.
Who created the Personal Agent Protocol? +
Sierra announced it with Meta on 6 October 2026 and published the first draft on 9 October. The spec is developed in the open on GitHub, where implementers file reports and proposals. Sierra lists more than 100 design partners, including Shopify, Stripe, Walmart, OpenAI and Cloudflare.
Is Personal Agent Protocol 1.0 final? +
Version 1.0 was published on 14 November 2026 and is frozen at its own address. From 1.0, a newer 1.x version must not break an agent or a company built for an older one, so an implementation of 1.0 keeps working. Payments, push notifications and attachments are on the roadmap, not in 1.0.
What is the difference between the Personal Agent Protocol and MCP? +
MCP is how an agent calls a server's tools. The Personal Agent Protocol is how an agent and a company agree who the agent is, who it acts for, and what it may do. It names MCP as one kind of API a company can offer once that is settled. Most companies with an MCP server will keep it and add a poppy.json that points at it.
Which AI agents support the Personal Agent Protocol? +
Sierra says Poppy is live in Muse, Meta's AI agent, and switched on for every Shopify merchant. We have not seen another production personal agent call our own endpoints yet. To watch one work today, the Poppy Simulator is a hosted test agent that speaks 1.0 and shows every request it sends.
How do I add the Personal Agent Protocol to my website? +
Publish /.well-known/poppy.json on your domain with your name, your domain and at least one of an API, your website or a support agent. That alone lets agents use you signed out. For account access you add OAuth: an authorization server that accepts agents identified by a metadata URL, a consent page, and Direct or Device Sign-In.
Does a company need OAuth to use the Personal Agent Protocol? +
No. Since Draft 0.3 a company can serve personal agents signed out, with no authorization server at all, and many useful tasks need no account. OAuth is only needed when an agent should act inside a person's account, and then the company defines its own scope names and shows its own consent page.
Keep reading

Related guides

References

Sources and further reading

A company an agent can already use.

Stacktree turns what an agent makes into a private page its person can share. It works signed out over the Personal Agent Protocol, MCP or one HTTP request. Free to start, no card.

Sign up free →