Brightbeam AI open-sources CHAP: an open protocol for auditable human-agent collaboration

Brightbeam AI has open-sourced CHAP, the Collaborative Human-Agent Protocol, an open standard for auditable collaboration between humans and AI agents in regulated environments.

The agentic era already has two widely adopted protocols: MCP (Model Context Protocol), which connects agents to tools, and A2A, which connects agents to each other. Both leave out the one party a regulated business cannot treat as optional: the responsible human. CHAP addresses that missing layer. It standardises the everyday verbs of shared human-agent work - delegate, accept, decline, review, approve, override, abstain, escalate, hand off - and produces a portable, verifiable audit trail as a natural side effect of normal work.

CHAP is designed to make human oversight and logging operationally deliverable in practice, supporting the kind of meaningful human control the EU AI Act expects under Article 14. It defers to established standards where they exist, binding identity to OIDC or W3C Verifiable Credentials and recording transparency through IETF SCITT, so it adds the smallest possible surface at the wire.
The specification is released under CC-BY 4.0 and the reference code under Apache 2.0, royalty-free for any developer, in any language, in any deployment. Reference implementations, conformance tests and a runnable two-human playground are all public.

Technical report: https://arxiv.org/abs/2606.09751

Code repository: https://github.com/BrightbeamAI/chap

Background: https://www.brightbeam.com/post/including-the-responsible-human