Personal Agent Protocol (Poppy): what it is, and what your website should publish
Poppy 1.0 gives personal AI agents a standard way to arrive at a business, say who they are and ask permission. Meta's Muse reads it and Shopify publishes it for every store. Here is how it works and what to publish first.
The short answer
Poppy matters because it is the first widely backed answer to a problem every business with a website now has. Agents are arriving, they look like customers, and until now there has been no agreed way for one to say "I am acting for this person, here is what I would like to do". It is also early. The first draft is six weeks old, the first consumer agent adopted it days ago, and most of what it promises depends on systems a small business does not have.
What is the Personal Agent Protocol?
The Personal Agent Protocol is a set of rules for how a personal AI agent, acting for one person, works with a company on that person's behalf. It covers four things: finding what the company offers, identifying the agent, getting the customer's permission for account access, and doing the task through an API, the website or the company's own agent.
Sierra, the customer service AI company co-founded by Bret Taylor and Clay Bavor, announced the idea with Meta on 6 October 2026. Its argument was simple. Personal agents mostly use websites the way people do, loading pages and clicking through forms, and when that fails they fall back to a phone line or a chat window. A direct, permissioned connection would finish the same task faster and with less risk (Sierra, 6 October 2026).
Version 1.0 arrived five weeks later, after testing with design partners against real websites, APIs and agents. It leans on standards that already exist: OAuth for sign-in, OpenAPI and the Model Context Protocol for APIs, and the Universal Commerce Protocol for shopping (Sierra, 13 November 2026). The specification is published under the Apache 2.0 licence (specification).
One caution on the version number. An independent tracker, poppy.md, counts four drafts between 9 October and 10 November and notes that no governing body has been named. The standard is moving quickly, and it is still run by the companies that wrote it.
How does Poppy work?
A business publishes a JSON file at /.well-known/poppy.json. A personal agent fetches it, confirms the file belongs to that domain, and learns what it may do: browse the site, call an API or talk to the company's own agent. For anything tied to an account, the customer signs in and approves the access first.
The specification describes the pieces in this order:
- Discovery. The file sits at a fixed address on the company's own domain and is served over HTTPS. The agent checks that the domain named inside the file matches the host that served it.
- Identity. Every request carries a session header that names the personal agent and the session. The specification also recommends that agents sign requests with Web Bot Auth, so a firewall or CDN can confirm the operator before the request reaches the site.
- Signed out or signed in. A signed-out agent can do what any visitor can: check stock, read a returns policy, ask a question. Account tasks need the customer to sign in through OAuth and approve named permissions.
- Permission hints. Each permission says whether it changes anything, and whether the change can be undone. If a hint is unclear, the agent must treat the action as irreversible.
- Channels. The file can point to APIs, to a conversation endpoint for the company's own agent, including a handover to a person, and to the website itself.
- Guidance and feedback. A company can list the tasks it will not handle through the protocol and suggest another route. Agents can report back when a task fails.
Two safety rules are worth knowing. Anything a company sends is treated by the agent as information, never as an instruction, which limits what a hostile page can make an agent do. And the specification admits its own limit on trust: the domain speaks for the company, so a lookalike domain can pass every check.
Who supports Poppy today?
On the agent side, Meta's Muse. On the business side, every Shopify store, according to Sierra, plus a small number of live partners. More than 100 companies are in the design partnership, but being a design partner is not the same as having shipped.
Meta Muse. Sierra says Poppy is on for everyone who uses Muse, which "automatically looks for Poppy first on every site it visits". That matters because of Muse's share. HUMAN Security's October benchmark, as reported by PPC Land, put Muse at 59% of the agent requests it saw. That is one vendor's network, counted in requests and not in people, but no other agent is close. Muse cannot be used in Australia today, and Meta had not announced a date here when it launched, ChannelNews reported. So for most Australian businesses, the agents reading this file today are acting for overseas customers.
Shopify. Sierra says every Shopify store now publishes a Poppy file automatically, and that Shopify co-wrote the commerce extension that connects Poppy to UCP. We tested the claim. On 16 November we requested the file from 10 online stores that also serve a UCP profile, a sign they run on Shopify's agent layer. Eight returned a valid file, including Allbirds, Brooklinen and ColourPop. Two returned a 404 for Poppy while still answering for UCP.
The eight files were near identical. Each declared protocol version 0.4, the release candidate, not 1.0. Each pointed agents to the store's UCP endpoint, offered one permission for preparing a checkout as a signed-in customer, and left the website section empty. In other words, Shopify's Poppy file today is mostly a signpost to UCP. Ten stores is a spot check, not a study, and it says nothing about whether an agent completed a purchase.
Live partners. Sierra names three. Mercury customers can have an agent check balances and start a payment, which Mercury tests against the account's approval policy. Turkish Airlines lets agents check fares and flight status and pull up a booking. Synchrony's work is described as exploratory.
Everyone else. The 60 partners added at 1.0 include Delta Air Lines, Hyatt, Intuit, Salesforce, Twilio, Wayfair, Wix and Zoom. OpenAI, Visa and Bank of America joined the design group in October. None of that tells you which agents other than Muse read the file today, and we found no statement from OpenAI, Google or Anthropic that theirs do.
What does a business have to publish?
At minimum, one file. Sierra's implementation kit says a company that offers only its website stops after publishing poppy.json without the sign-in section. Everything beyond that, such as APIs, a company agent and customer sign-in, is optional and added in stages.
| Stage | What you publish | What a personal agent can do | Who it suits |
|---|---|---|---|
| No file | Nothing | Browse the site as a signed-out visitor, if your firewall lets it in | Most websites today |
| Website only | poppy.json with your name and domain, no sign-in section | Confirm which business the domain speaks for, then browse the site signed out | Lead generation and brochure sites |
| Interfaces | The file plus an API description or a company agent endpoint | Ask questions and fetch answers without loading pages | Businesses with a public API or a support agent |
| Sign-in | OAuth sign-in and named permissions | Act on a customer's account once the customer approves | Banks, airlines, subscriptions, member portals |
| Extensions | An industry extension, such as commerce | Shop as a guest or a signed-in customer through UCP | Online stores, mostly through their platform |
Source: Personal Agent Protocol specification , Sierra implementation kit
The kit itself is built for developers. It is a skill that a coding agent follows, with a command-line validator that acts as a test personal agent against your deployment. Sierra recommends running it against a test environment, because it opens real sessions.
Where the file comes from depends on your platform.
- Shopify. Published for you, as described above. Open yourstore.com/.well-known/poppy.json to confirm.
- WordPress and WooCommerce. No official support that we could find. Two independent plugins in the WordPress directory serve the file: Agent Adapt, which says it implements draft 0.2, and Allowfy for WooCommerce, which fetches the file from the vendor's servers. Both are young, and neither is endorsed by the protocol's authors. The file itself is small enough to serve as a static route. The harder part is everything around it, which is why a WordPress specialist such as WP Creative would check the security plugin, the host and the CDN together, so the file reaches agents and not only browsers.
- Wix. Wix is a design partner. We found no date for when Wix sites will publish the file.
- Custom sites. Any developer can add a static JSON file at the well-known path.
How is Poppy different from UCP, WebMCP and Web Bot Auth?
They answer different questions. Poppy is about arrival and permission. UCP is about carts and checkout. WebMCP is about actions on a page. Web Bot Auth is about proving which operator sent a request. A site can use all four, and Poppy is designed to point to the others.
| Standard | The question it answers | What the website publishes |
|---|---|---|
| Personal Agent Protocol (Poppy) | Who is this agent acting for, and what may it do here? | A file at /.well-known/poppy.json |
| Universal Commerce Protocol | How does an agent build a cart and check out? | A profile at /.well-known/ucp |
| WebMCP | Which actions on this page can an agent call directly? | Tools declared in the page's own code |
| Web Bot Auth | Did this request really come from the operator it claims? | Nothing; the CDN or firewall verifies a signature |
Source: Personal Agent Protocol specification , UCP specification, Cloudflare
Sierra describes the split with UCP directly: UCP handles finding products, filling carts and checkout, and Poppy handles the step before, when an agent arrives and finds out what it is allowed to do. Our guide to the Universal Commerce Protocol covers the shopping half, and our guide to WebMCP covers forms and on-page actions.
Payment is a separate layer again. The protocol's roadmap lists payments as an open topic, without a date. Today an agent that pays does so with something like ChatGPT Wallet, through your existing checkout.
What happens if your website has no Poppy file?
Nothing breaks. The agent treats your site as it would have a month ago and browses it as an ordinary signed-out visitor. That is the specification's own fallback, so the quality of the website still decides the outcome.
This is the part most coverage skips. Three points follow from it.
- A Poppy file does not open the door. If your firewall challenges or blocks automated visitors, the agent may never reach the file. Poppy recommends that agents sign their requests so bot defences can recognise them, but your defences have to be set up to accept that. Our guide to AI agent detection explains how that recognition works.
- Opting out is a request, not a lock. The file can ask agents not to browse the website. The specification is plain that this states a preference and cannot be enforced.
- The file cannot fix the page. An agent sent to a contact form that only works with a mouse, or to a price that loads late through a script, fails in the same way with or without Poppy.
Our view: for a business that generates enquiries, Poppy changes little this quarter and a great deal over two years. Today its value is identity. A website-only file tells Muse which business stands behind your domain. The larger prize comes when an agent can ask your business a question and get a structured answer back, and that needs systems most small businesses will get from their platform, not build themselves.
Where is this heading?
These are forecasts, not facts.
- Platforms will publish the file, as Shopify has. We expect the large site builders and the main WordPress ecommerce plugins to ship Poppy support during 2027. Most businesses will get it as a setting.
- More agents will read it. Muse is first. With OpenAI in the design group, we expect at least one more major agent to check for the file within a year.
- Identity will reach the firewall. Once agents arrive signed and named, security tools can treat a known personal agent differently from an anonymous bot. Blocking everything automated will start to cost real customers.
- The standard will change. Four drafts in five weeks, and Shopify's own files a version behind, tell you what to expect. Anything hand-built now will need maintaining.
- Service businesses get a version of this. Bookings, quotes and account questions fit the model as well as shopping does. A clinic or a trades business will not build an API, but its booking system might.
The risk today is not being late to Poppy. It is spending a month on an agent API while the website an agent actually visits still hides its phone number behind a script.
What to do next
- Check whether you already have a file. Open yourdomain.com/.well-known/poppy.json in a browser. On Shopify you should see JSON. Anywhere else, a 404 is normal today.
- Make sure agents can reach the site at all. Run the free AI crawler access checker to see which AI visitors your robots.txt and firewall turn away.
- Test the page an agent would land on. Run the agent readiness check on your main service page and your contact page. If the content is missing without scripts, fix that before anything else.
- Decide what a signed-out agent should be able to do. Write it down in plain words: read prices, check availability, send an enquiry, book a call. That list is your brief for the file and for the website.
- Publish the basic file if it is cheap to do. Ask your developer to serve a website-only poppy.json with your business name and domain. Skip sign-in and APIs unless you already run them.
- Ask your platform and your booking or CRM vendor. One question: do you publish a Poppy file for customers, or will you? Their answer matters more than anything you could build this year.
- Review it in three months. The specification is changing quickly. Put a date in the diary to recheck the version and which agents read it.
Frequently asked questions
Is the Personal Agent Protocol the same as MCP?
No. The Model Context Protocol connects an AI model to tools and data. The Personal Agent Protocol sits above it and covers how one person's agent finds a business, identifies itself and gets that person's permission. A Poppy file can list an MCP server as one of the ways an agent may reach the business, alongside an OpenAPI description or the website.
Do I need a Poppy file if I am on Shopify?
You should already have one. Sierra says Shopify publishes the file for every merchant, and eight of the ten stores we checked on 16 November served it. Open yourstore.com/.well-known/poppy.json to confirm. If you see a 404, ask Shopify support, particularly if your store uses a custom front end.
Does a Poppy file let AI agents log in to customer accounts?
Not by itself. Account access needs the business to offer OAuth sign-in, and the customer must sign in and approve each named permission. A file with no sign-in section allows only what any signed-out visitor can do. The customer can withdraw access later, and the business chooses which permissions exist in the first place.
Can I use Poppy to block AI agents from my website?
No. The file can state that you would prefer agents not to browse the site, but the specification says that is a preference and not a promise. Blocking is still done at the firewall or CDN. If you do block, decide which agents and why, because a personal agent is usually a customer trying to do something.
Can an AI agent use your website?
Check it free.
PostCliq checks whether AI agents can read your pages, whether visitors can reach your contact path, and puts three fixes first.
Free. No credit card.
