An agent should not stop at chat
Most agents begin in a conversation. That is useful while the work is still being shaped, but many agents eventually settle into a clear job: collect a known set of inputs, make a bounded decision, and return a predictable result. At that point, a chat window is no longer the best interface for every user.
Agent Apps is our attempt to give that job a proper surface. In Agentlas Site, you select an agent, team, firm, or group you already own. The app is then designed around that source instead of being rebuilt as a disconnected frontend with a second set of instructions hidden behind it.
One source, one runtime boundary
The important piece is not the page template. It is the contract between the interface and the agent. Agentlas pins the selected source, input and output shapes, and allowed capabilities before the app runs. A local preview talks to a capability-scoped loopback runtime, not an open-ended process with access to the whole machine.
That gives the interface room to become specific without making its authority vague. A research agent can ask for a topic and evidence threshold. A support team can expose intake fields and a review result. The surface changes; the ownership and execution contract remain visible.
From private tool to public surface
The current publishing path is being designed around three fixed providers: Vercel, Railway, and Render. Publishing is an explicit action, and each provider has its own consent step. Provider tokens stay in the operating system Keychain rather than being written into the generated app.
The deployment boundary also stays honest. Agentlas can prepare and publish the application, but the provider account, billing, terms, and service availability still belong to the user. For Render, the first service can be created without sending an LLM key at all.
The boundary is part of the feature
It would be easy to describe this as an agent-to-app button. That misses the work that makes the button trustworthy. A useful Agent App needs a stable contract, a narrow local runtime, clear provider consent, and a result that can be traced back to the agent the owner selected.
- The source agent remains identifiable.
- Inputs, outputs, and capabilities are frozen before execution.
- Publishing requires an explicit provider choice and consent.
- Credentials remain outside the generated interface.
Where it stands
Agent Apps is in active development inside the current Agentlas Desktop workstream. The selection flow, local runtime boundary, preview surface, thumbnails, and provider publishing contracts exist in code, but this is a first look, not a general-availability announcement. We will publish final availability and setup details after the end-to-end release path clears verification.
