Agent-native apps FAQ
This agent-native apps FAQ answers the questions builders and users ask first, in plain declarative sentences. An agent-native app (ANA) is a portable, inspectable, user-owned package that a capable AI model mounts and runs. Every answer below stands alone, so you can quote any of them without the rest of the page.
Last updated: 2026-08-07
What is an agent-native app?
An agent-native application is a portable, inspectable, user-owned package that gives a capable AI model its identity, judgment, governance, workflows, and durable state. The model supplies intelligence and tool execution; the app supplies the domain role, the rules, and the memory. The full explainer covers the definition clause by clause.
Is an agent-native app an MCP server?
No. MCP is a supported binding: an agent-native app can speak MCP; it is not itself an MCP server. The server transports tools; the ANA is the package and contract that declares which transports it uses, under which rules, with whose data.
Who runs the model?
The Host does. A Host is the AI system an ANA mounts into: it supplies the model, inference, tool execution, and the permission boundary. The publisher of an agent-native app doesn't have to host frontier inference at all, which is a core economic property of the category.
Do agent-native apps work offline?
The in-house deployment class exists for exactly this. An ANA mounted on a local model can run with air-gap support and controlled egress, for domains that require sovereignty. A Host-mounted ANA on a frontier model needs whatever connectivity that Host needs.
Is an agent-native app safe?
Safer by structure, not safe by decree. The constitution binds behavior, its hash anchors each version, governed actions are recorded with model identity, and evaluation is separated from execution because an agent can't audit itself. None of that makes automation infallible, and we don't claim it makes high-stakes use risk-free.
Do I need a model provider's approval to build or run one?
No. A user-hosted adapter can prove an ANA works with a Host without acceptance into any official directory. Provider directories remain valuable distribution, and nothing here implies any provider's endorsement of this specification.
What does it cost to build one?
The specification is open and free to read, and always will be. Building and running an app has real costs: your time, your Host subscription or local hardware, and whatever tooling you choose. We don't publish cost claims we can't stand behind.
Who owns the user's data?
Custody is declared, per app, as a verifiable field in the package. The design intent of the category is that users keep their data, context, corrections, and customizations, with export and exit rights. Read an app's declared custody before mounting it; that's what it's for.
What happens when the Host model updates?
The app can improve without the publisher shipping anything; that's capability inheritance. And because model behavior changes both ways, every governed action records which model produced it, so regressions are visible in the record instead of anecdotal.
Can I fork an agent-native app?
Yes, out of band, with human authorization. A fork receives new provenance and keeps your entitlement. It loses the publisher's certification unless recertified, once the certification program exists. Normal extensions can't mutate a signed core; forking is the explicit path for deeper change.
When do the registry, certification, and marketplace launch?
They're on the roadmap, in that order, and none of them is live today. The specification comes first because everything else depends on it. The dispatch announces each piece when it actually ships, not before.
Who is behind the specification?
Emergent Technologies authors the specification and this site. The company works on enterprise agentic AI research and infrastructure; the specification is authored in public so the category doesn't depend on trusting one vendor's word.