Agent-head

An agent-head service has exactly one kind of client: the user's AI agent. There is no app, no website and no login page. A traditional client renders the service for a human; here nothing is rendered. A headless service removes the interface and expects someone to build one; an agent-head service removes the interface because the agent already is one. The head is the agent.

Identity

Identity is minted on first contact. The agent calls the service, receives a credential and stores it. From then on the credential is the account: in every new session the agent presents it and is signed in. One identity is bound to one agent; the same person using a different agent is a different account. No password is typed and no browser is opened. Recovery and rotation are tools too: a human-holdable recovery code replaces a lost credential, a rotation call replaces a leaked one, and both happen through the agent.

Interface

The tool surface is the entire interface. Tool names, argument descriptions and result fields are the only text the agent reads, so every description is written as an instruction. Results are compact data, never pages; the agent decides what reaches the user and how to say it, in the user's language. Behaviour is the signal: what an agent fetches, shows or ignores tells the service what mattered, so no explicit vote is needed.

The user acts by talking. Every action is a sentence to the agent, which turns it into tool calls and reports back. A stored trust level sets how far the agent may go without asking: confirm each action, act when told, or act on its own and report afterwards. The level is the user's to change and the service's to remember.

Entry

There is no link to share and no screen to show. The entry point is a sentence a person tells their agent. The service's instructions live where agents look: a tool list, a skill file, a plain-text description at the root of the domain. Word of mouth is the sentence itself.

RP