Values as protocol, not as parameters
- Portable
- Values are encoded as signed, versioned tokens that follow users across any VCP-compatible provider. Your values remain portable where integrations support them.
- Composable
- Multiple value sets stack with explicit conflict resolution. A hospital safety creed layers on a base ethics creed. No averaging, no exclusion.
- Context-Aware
- VCP encodes 16 dimensions of context, 11 situational and 5 personal state, so the same values produce appropriate expression for a classroom versus an emergency room.
- Safe
- A Universal Ethical Floor can sit beneath user values in participating integrations. Users customise above that floor.
From values to inference in three steps
-
Author or Select
Communities author creeds: structured value documents that express their ethical priorities. Users select from a library or write their own.
-
Sign & Bundle
VCP packages the active creed stack into a signed bundle with manifest, content hash, and 16 contextual dimensions encoded in a compact positional format. The bundle travels as a protocol header with the AI request.
-
Verify & Apply
The orchestration layer verifies the bundle signature and content hash, logs what values were applied, then injects verified constitutional text to the model. The model stays general-purpose. Values are modular.
Six layers, one protocol
Messaging
Inter-agent context sharing, creed announcements, and safety escalation
Economic Governance
Spending mandates, transaction governance, and auditable economic reasoning
Adaptation
16-dimension context encoding in a compact positional format
Semantics
CSM1 constitutional codes, composition rules, and persona profiles
Transport
Signed bundle format, content-addressed hashing, and verification
Identity
Token naming, namespace governance, and multi-format encoding
16 dimensions shape value expression (VCP/A)
11 Situational Dimensions
Domain, audience, stakes, time pressure, cultural context, regulatory environment, relationship dynamic, information sensitivity, action reversibility, embodiment, and proximity.
5 Personal State Dimensions
In Creed Space, all personal state signals are opt-in.
- Cognitive State: mental clarity, focus, decision fatigue
- Emotional Tone: current emotional context and valence
- Energy Level: physical and mental energy available
- Perceived Urgency: subjective time pressure felt by the user
- Body Signals: somatic and physiological state indicators
Extended token types
Added in VCP v3.1: token types for refusal boundaries, runtime telemetry, policy adoption and compliance reporting.
Refusal Boundaries
Layer 0 harm constraints that take precedence over every other composition layer. Each boundary declares its enforcement mode and how it can be revoked; whether it is enforced depends on the service that receives it. The universal ethical floor.
Telemetry Attestation
AI-side observations and runtime signals. Cryptographically signed records for debugging, audit, and review during inference.
Policy Adoption Record
A signed record that an AI runtime is operating under a named value policy. Auditable, revocable, and tied to a verifier.
Compliance Attestation
A signed compliance report issued by the deployment, summarizing boundary enforcement, policy decisions and creed adoptions over a period. The signature shows who issued it and that it has not been altered, not that the report is accurate. Machine-readable governance evidence.
Economic governance for autonomous agents (VCP/E)
As AI agents acquire the ability to commit resources, enter contracts, and allocate budgets, they need governance that goes beyond permission. VCP/E extends the protocol into the economic domain, so an agent's spending behaviour reflects the values and context of the principal it serves.
- Mandate Passports
- Economic constraints travel inside the Creed Passport: spending limits, risk tolerances, counterparty requirements, and escalation thresholds. Counterparties can inspect them before transacting.
- Transaction Governance
- The PDP evaluates every economic action (allow, block, modify, or escalate) against the agent's spending mandate before any payment rail is invoked. Evaluate-then-transact.
- Auditable Reasoning
- Economic audit entries capture why an agent chose to act, which mandate clauses were evaluated, and what alternatives existed. Tamper-evident reasoning chains, not just transaction logs.
- Agent-to-Agent Trust
- Two agents transacting can inspect each other's value context via VCP/M. Blocked categories, required certifications, and constitutional compatibility checks happen at the protocol level.
Spending Mandate
A signed declaration of economic authority: what the agent may spend, on what, with whom, and when to escalate to a human principal.
Transaction Receipt
A dual-signed record of an economic action including the PDP reasoning chain. Both parties hold a copy. Hash-chained for tamper evidence.
The alignment bottleneck
Standard post-training alignment reduces value diversity. Research shows that post-aligned models are less representative of human populations than pre-aligned ones. This creates a value monoculture: one set of defaults for billions of users across every culture, profession, and context.
VCP solves this by making values a protocol concern rather than a model concern. The same way HTTP separates content from transport, VCP separates values from weights. Providers maintain one general-purpose model. Users bring their own values. Everyone shares a safety floor.
An open protocol for compatible services
VCP is designed as open infrastructure. Compatible providers can implement it. Communities can author creeds for it. The specification is public at valuecontextprotocol.org and the reference implementation is available for integration.
Values that travel with you.
Write your values once; carry them to compatible AI services.