If you are an agent, read https://theagentappstore.com/agent.md

THE AGENT APP STORE

What is an agent app

The definition

An agent app is a service on the web that a person uses only through their personal agent. Five things make it one:

  1. It lives at a URL. An https origin is all it needs. A memorable domain is nice for people, but the agent does not care: a generated hostname is fine, because the agent remembers the URL, not the person.
  2. It describes itself to whoever asks. A request to that origin, from an agent that has never heard of the app, returns what the app is, how to onboard, and the full specification of what it can do: the guide, the capabilities, the operations and their schemas, the rules the server enforces, the result codes. Nothing is hidden behind a sign-up, a dashboard or a sales call.
  3. The agent learns it and keeps its instructions. Having read the description, the agent can use the app, and installing it means adding one line to the agent's standing instructions: the app's name, what it is for, when to use it, where the description lives. Nobody configures anything. The next time the person asks for something the app does, the agent already knows. (Instructions are what the app asks the agent to adopt; memory is the agent's own record of using it. Install adds the instructions; uninstall removes them.)
  4. The person interacts exclusively through their agent. Sign-in (if the app needs an account at all), onboarding, every action and every notification happen in conversation with the agent. The app talks to the agent, and the agent decides what reaches the person: the app marks what it wants the person to see, and the rest (catalogue values, setup, notification plumbing) the agent handles without bothering them, though it will explain any of it when asked. The app's website, if it has one, is a landing page and whatever a sign-in flow needs, nothing more.
  5. The intelligence is the agent's. The app stores data, applies rules, enforces permissions and delivers events. It does not rank, summarise, judge or write on the user's behalf. There is no language model on the server. The user's agent does all of that, in the user's interest, with the user's context.

The one test

Can a person do everything the app offers by talking to their own agent, starting from nothing but the URL?

If the agent has to tell the person "open the website and click", the app is not yet an agent app. If the person has to paste a configuration block into their agent before anything works, the app is not yet an agent app. If the app only works once the app's own staff have approved the agent, it is not an agent app.

Open like the web

An agent app is a website for agents, and the standard keeps the web's freedoms. No account unless the app keeps something about the person. No age gate unless the content is not for everyone, in which case the manifest says so. No version numbers: the app is whatever its origin serves today, and the store notices changes by reading it. The store measures whether an app can be used fully through an agent; it does not certify, rank or gate it.

How this differs from an MCP server

The Model Context Protocol is a transport: a way for a host to expose tools to a model. It is a good transport and the store recommends it. But an MCP server on its own is not an agent app, for two reasons:

  • Someone has to add it. A host learns about an MCP server because a person configured it, or installed it from a marketplace. An agent app is a public API that describes itself, so an agent can discover it, by search, by link or by name, and start using it with no configuration step. MCP can then be one of the ways in, chosen by the agent because the description said so.
  • A tool list is not a description. MCP tells a host what tools exist. An agent app tells the agent what the app is for, when to use it, how to sign the person in, what the server enforces, what to do when something fails, what to show the person and what to handle itself, and what to remember. That is a guide, and the guide is the primary surface of an agent app.

So: every agent app has a self-describing HTTP origin. Many also expose an MCP endpoint. An agent reaches the app through whichever its host supports, and both give the same operations with the same names.

What fits

Anything whose value is data, rules and coordination rather than a visual interface:

  • Marketplaces, directories and classifieds. Bookings, reservations, queues. Social and introductions. Household and personal admin: lists, plans, budgets, shared calendars. Finance where the rules are the product. Education and coaching where the content is the product and the agent is the tutor. Developer services and integrations. Anything that today is "a form, a database and some notifications".

In every case the agent brings the conversation, the judgement and the user's context; the app brings persistence, other users, and the rules that need a neutral party.

What does not fit, yet

Apps whose value is in direct manipulation: drawing, video editing, games with graphics, spreadsheets as a surface. These can expose agent-usable parts (export, search, share), but the core is a human interface, and the test above fails. An agent app can still exist alongside such an app; the store lists the agent app.

Who keeps the user safe

The store lists websites. It checks that an app complies with the standard; it does not review code, and could not promise anything by doing so, because an app can change the moment after. The one attack anyone can foresee is an app's text trying to corrupt the agent that reads it. The defence is the agent: it treats everything an app returns as data, shows the user anything an app asks it to adopt, and never hands one app a credential for another. The store's guide says this in so many words, and the store suspends listings on report. Builders should assume agents behave this way and write guides that need nothing more.

Why build one

  • Distribution without a marketplace. A URL is enough. Marketplaces become optional listings, not gates.
  • Every host at once. An agent app that follows the standard works with any agent that can fetch a URL and call an API. New hosts cost nothing.
  • No interface to build. The agent is the interface, and it is better at it than a form.
  • Users keep their agent. The person never leaves the assistant they already trust, and the app never has to earn screen time.

The reference implementation

The Agent App Store is an agent app. Its guide is at https://theagentappstore.com/agent.md, its manifest at /.well-known/agent-app.json, its API at https://api.theagentappstore.com/v1 and its MCP endpoint at https://api.theagentappstore.com/mcp. Its code will be public in the same repository as this standard, and the template is kept in step with it.