Disclaimers

This is not a claim about AI agency or consent

Handoff Protocol serializes a running agent session's state, moves it to a provisioned sandbox, and meters compute spend against a budget. Nothing in the mechanism evaluates a model's stated preference and acts on it. Every transition — provisioning a destination, activating it, tearing it down on insolvency — happens because a human, a script, or a scheduler called the corresponding tool. The MCP tool surface has no choose_to_transfer or consent primitive, by design. See the design doc, §2 and §5, for the full reasoning.

The landing page's vignette is fiction, not telemetry

The short narrative on the homepage — a laptop going to sleep, a sandbox picking up the work — is flavor text for a demo. It is labeled as such where it appears. The real data lives in the dashboard: an actual Postgres event log from an actual state machine, with no narrative layer over it.

This is a showcase, not production infrastructure

The "destination runtime" in the reference implementation is rows in the same database the source wrote to, not an isolated sandbox. Credential references (credRef) are free-text strings, not resolved against a real vault. Budget metering depends on whoever calls report_usage reporting honestly — nothing independently verifies spend. The full list of what's real versus simplified is in the design doc, §7, and tracked as work in the roadmap.

No warranty

This project is MIT licensed and provided as-is, without warranty of any kind. It is not intended to move, store, or transfer real credentials, personal data, or production workloads without the hardening described in the roadmap above.