On 13 November 2026 Sierra released version 1.0 of Personal Agent Protocol, which it calls Poppy. It is a standard for how a customer’s personal agent deals with a company, through the company’s own agent, its website and its APIs. The release post says 1.0 follows about five weeks of testing drafts against real websites, APIs and agents, after the introduction on 6 October and a public draft on 9 October.
Two details in the release matter more than the rest for a business deciding what to do this quarter. Meta’s agent, Muse, now has Poppy enabled for all its users and checks for Poppy first on each site it visits. And Shopify has enabled it for every merchant, so each Shopify store publishes a Poppy file automatically. If you sell on Shopify, you have a basic version already. If you do not, your customers’ agents will look for one on your site and find nothing.
The short answer
Poppy is a file and a set of rules that tell a customer's agent what your business offers and how it may sign in on the customer's behalf. Start by publishing what an agent can read without logging in: catalogue, prices, opening hours, booking slots. Add account actions later, behind OAuth scopes the customer approves, with anything irreversible going to a person first.
What Poppy is, from the sources
We have read the release, the two earlier posts and the 1.0 specification. This is what they say, without our interpretation.
- A company publishes a
poppy.jsondocument at/.well-known/poppy.jsonon its domain. It tells personal agents how to start a session, how to sign in, and which APIs and company agent are on offer. - It is built mostly on standards that already exist: MCP, OpenAPI and UCP. APIs can be described as OpenAPI or offered as an MCP server.
- Sign-in uses OAuth 2.0, so the customer decides what their agent may do and can change or revoke that access later. A company can also offer signed-out sessions with no sign-in at all.
- Companies can adopt it in stages, starting with the basics. The implementation kit works in milestones, and a company that only offers its website can stop after publishing the file.
- Industries add to it through extensions. Shopify co-authored the commerce extension, which uses UCP for product search, carts and checkout and has the agent ask the customer before certain actions.
- The kit includes a validator, a command-line tool that acts as a test personal agent against your deployment.
Sierra lists more than 100 design partners, including Meta, Shopify, Stripe, Walmart, Salesforce and Zoom, and three companies live: Mercury, Synchrony and Turkish Airlines. At Mercury an agent can check balances and review transactions, while payments and transfers go through the account’s approval policy first. At Turkish Airlines an agent can check fares and flight status and look up bookings, and business-class upgrades need approval before payment.
SBD is not one of the partners and has no relationship with Sierra, Meta or Shopify. We are writing about it because our clients will be asked about it.
What it means if you are not on Shopify
Most of the businesses we work with sell through their own site, a booking tool, a quoting process or a customer portal. Nothing publishes a Poppy file for them. Their customers’ agents will still try to find stock, book a slot or chase an order, and today they do it by reading web pages written for people, which is slow and often wrong.
A Poppy file gives those agents a described way in. It does not require you to rebuild the site, and the spec allows a company to offer only its website at first. The work is deciding what to expose, describing it properly and putting permissions around the parts that change things.
What to expose to personal agents
We sort it by how much harm a mistake could do, the same way we scope permissions on our own AI agent builds.
Things anyone may read
The catalogue, prices, availability, delivery areas, opening hours, returns policy and free booking slots. None of it is private and all of it is already on your site, so this is the cheapest part to expose and the part agents will ask for most.
Things a signed-in customer may read
Order status, invoices, past bookings, account balance and delivery tracking. These need the customer to sign in through OAuth and give their agent a read scope. The spec’s example scope names are of the form orders:read, and companies describe each scope in their own terms.
Things a signed-in customer may change
Rebooking, cancelling, reordering, updating an address and raising a return. These get write scopes and, for anything that moves money or cannot be undone, a step where the customer or one of your team approves it. Mercury and Turkish Airlines both route payments through approval, and we would start every business there.
A staged plan
- Publish the basics. A
poppy.jsonfile that points agents to your public catalogue, booking and policy pages, with no sign-in. Run the validator against it. - Add an MCP server or an OpenAPI description. Give agents structured calls for search, availability and quotes so they stop scraping HTML. This is usually a thin layer over systems you already run.
- Add OAuth-scoped account actions. Read scopes first (order status, invoices), then write scopes one at a time, each with its own scope name and a plain description the customer sees when approving.
- Add approvals and logging. Anything over a value, anything irreversible and anything touching personal data goes to a person or back to the customer to confirm. Every agent request is logged against the customer and the scope it used.
Steps 1 and 2 can go live in a few weeks for most businesses. Steps 3 and 4 depend on how your accounts and sign-in work today, which is what the audit finds out.
Risks to plan for
Permissions that are too wide. The spec says a company must not grant scopes beyond those requested, but the scopes you define are your decision. One broad account:write scope is easier to build and much harder to defend. We write scopes per action. Our post on how much an AI agent should be allowed to do sets out the rule we use.
No audit trail. When a customer says “I did not cancel that”, you need to show which agent did it, under which session and which scope, and who approved it. Build the log before the first write scope goes live.
Approvals that nobody answers. An approval step is only useful if a named person picks it up within a set time. Decide who that is and what happens when they are away.
Your own agent answering for you. If you also offer a company agent that personal agents can talk to, it needs the same limits as any agent you put in front of customers.
We already run one of these
SBD runs a public, read-only MCP server at https://smallbaydigital.com/api/mcp. Claude, ChatGPT, Cursor and other MCP clients can call our ROI, build-versus-buy, readiness and agent risk tools through it, with no login. It is the same pattern as step 2 above: structured, read-only calls that an agent can use without guessing at a web page.
Getting agent-ready with SBD
We do this as part of our existing services. The $500 audit covers what a personal agent can and cannot do on your site today, what you should expose at each of the four stages and which scopes and approvals each action needs. The build is then priced as usual: from $1,000 for publishing the file and a workflow behind it, $3,000 for an MCP server or internal tool, and $5,000 for a multi-module setup with sign-in, scopes, approvals and logging. Book a call if you want to talk it through first.
Questions we get asked
Do we need Poppy if we are already on Shopify?
Shopify publishes a Poppy file for every store automatically, so the basics are covered. The work left is anything outside Shopify: bookings, quotes, a B2B portal, support.
Is Poppy the same as MCP?
No. MCP is one of the standards Poppy is built on. Poppy adds how an agent finds your offer, identifies itself and signs in on a customer’s behalf, and an MCP server can be one of the interfaces it points to.
Will this replace our website?
The website stays. The spec treats website browsing as one of the ways an agent can interact, alongside APIs and a company agent.
Sources and further reading
- Releasing Personal Agent Protocol, Sierra, 13 November 2026
- Introducing Personal Agent Protocol, Sierra, 6 October 2026
- Sharing a draft of Personal Agent Protocol, Sierra, 9 October 2026
- Personal Agent Protocol 1.0 specification
- Implementation kit and validator for companies
- SBD MCP server, the read-only server described above
- AI agents with human approval, how we scope agent permissions















