CHAP – Collaborative Human-Agent Protocol
An open protocol for auditable human-agent collaboration, giving approvals, overrides, handoffs, and escalations a structured, verifiable, hash-linked record compatible with MCP and A2A.
At a Glance
About CHAP – Collaborative Human-Agent Protocol
CHAP (Collaborative Human-Agent Protocol) is an open standard published by Brightbeam AI for structuring and auditing the moment a human steps into an AI agent workflow. It captures approvals, overrides, handoffs, and escalations as typed, hash-linked envelopes rather than letting them dissolve into chat threads and ticket comments. The specification is licensed CC-BY 4.0 and the reference code is Apache 2.0, royalty-free for any language or deployment.
What It Is
CHAP defines a JSON-RPC 2.0 wire protocol with seven core methods — workspace.describe, participant.join, participant.leave, task.create, task.update, task.complete, and audit.read — that together create a shared, policy-bound workspace where humans and agents collaborate. Every action produces an envelope whose prev_hash field chains it to the previous entry using SHA-256 over JCS-canonicalised content, making the log append-only and tamper-evident. Optional profiles layer on top of the small core: review/1.0 adds approve, reject, override, abstain, and escalate; security-signed/1.0 adds Ed25519 per-message signatures; audit-scitt/1.0 anchors the chain in an IETF SCITT transparency log for offline-verifiable receipts.
How the Override Envelope Works
The decide.override envelope is the centrepiece of CHAP. When a human edits an agent's draft, the envelope captures:
diff— an RFC 6902 JSON Patch describing exactly what changedrationale— a free-text reason typed by the reviewerintent_preserved— a boolean distinguishing a refining override (same decision, better expression) from a substituting override (different decision entirely)tags— a controlled vocabulary the team defines, which accumulates into a supervision dataset as a side-effect of normal reviewpolicy_refs— references to the policies that governed the decision
The distinction between refining and substituting overrides matters operationally: a high refining rate on a policy clause points to weak retrieval or a poor template; a high substituting rate points to ambiguous policy or missing task context. These tune to different fixes.
Protocol Composition and Standards Reuse
CHAP is designed to sit beside MCP and A2A rather than replace them. MCP tool calls are recorded as citations inside CHAP artefacts; A2A peers appear as bridge participants in the workspace. The protocol defers to established standards throughout: JSON-RPC 2.0 for the wire, OIDC and W3C Verifiable Credentials for identity, IETF SCITT for transparency logs, and in-toto for attestations. The repository ships ready-to-adapt integration guides for CHAP + MCP, CHAP + A2A, and CHAP + OIDC/OAuth2, plus five framework bridges covering LangGraph, Pydantic AI, AG2, LlamaIndex Workflows, and Google ADK.
Deployment and Setup Path
Reference implementations exist in both TypeScript (npm: @brightbeamai/chap-coordinator) and Python (PyPI: chap-coordinator). Both emit identical wire bytes; the audit chain is byte-for-byte the same regardless of which client made the call. A five-minute start guide runs the reference server locally with SQLite persistence and sends envelopes with curl — no SDK, identity provider, or signing setup required for the first walkthrough. The conformance harness ships in the repository and the GitHub README reports 23/23 conformance vectors passing as of the latest release.
Update: v0.2.11
The latest release is v0.2.11 (published 2026-08-20), described in the GitHub metadata as "coordinator-mcp namespace casing." CHAP is currently at v0.2 — a public draft. The homepage states the wire format and schemas are stable for review; the Core runs; the review profile and a routing-aware playground are runnable; and conformance scaffolding is in the repository. Wire-format changes in 0.x remain possible. The headline step to 1.0 is a second independent implementation — independent meaning from outside Brightbeam; the two existing reference implementations (TypeScript and Python) are both from Brightbeam and do not count toward that milestone. Hierarchical workspaces are an explicit candidate for v0.3, and hybrid post-quantum signatures are named as future work.
Governance and Open-Source Model
The specification is governed to resist capture: at least three Steering Committee seats must be held by people not employed by the largest contributor, and no single party — including Brightbeam — can steer it unilaterally. A CHAP Enhancement Proposal (CEP) cannot be accepted without a working reference implementation, and vendor-specific extensions are explicitly out of scope. The project cites the pattern of TCP/IP, HTTP, OAuth, OIDC, and MCP as evidence that open shared standards become the basis of many platforms, and frames CHAP as the accountability layer that human-agent collaboration will need to standardise one way or another.
Community Discussions
Be the first to start a conversation about CHAP – Collaborative Human-Agent Protocol
Share your experience with CHAP – Collaborative Human-Agent Protocol, ask questions, or help others learn from your insights.
Pricing
Open Source
Fully open-source under Apache 2.0 (code) and CC-BY 4.0 (specification). Free to use, modify, and deploy in any language or environment.
- Seven core JSON-RPC 2.0 methods
- TypeScript and Python reference implementations
- Conformance harness with 23/23 test vectors
- All optional profiles (review, security-signed, audit-scitt, etc.)
- Five framework bridges
Capabilities
Key Features
- Seven core JSON-RPC 2.0 methods for workspace, participant, task, and audit management
- Append-only, SHA-256 hash-linked evidence log
- Structured override envelopes with diff, rationale, intent_preserved, tags, and policy_refs
- Optional profiles: review, modes, routing, deliberation, whisper, handoff, identity-oidc, identity-vc, security-signed, audit-scitt, control
- Ed25519 per-message signatures with JCS canonicalisation
- IETF SCITT transparency log anchoring for offline-verifiable receipts
- MCP tool calls recorded as citations inside artefacts
- A2A peers appear as bridge participants
- OIDC and W3C Verifiable Credentials identity binding
- Override analytics and supervision dataset accumulation as a side-effect of normal review
- Conformance harness with 23/23 test vectors passing
- TypeScript and Python reference implementations with SQLite persistence
- Five framework bridges: LangGraph, Pydantic AI, AG2, LlamaIndex Workflows, Google ADK
- 12 worked scenarios from solo developer to GMP-regulated manufacturing
- Five-minute curl-only quickstart with no external dependencies
Integrations
Demo Video

