How another agent asks a person for something.

252 scopes, five tiers, 8 purposes.

The vocabulary is generated, versioned, and served as a single public document at /pchp/scopes.json. Fetch it, cache it against its version, and build against the names — not against prose on a page, including this one.

What may EVER happen to a field.

The tier is not advice. It is the rule that governs a scope for the whole of its life, and it is the part an integrating agent most needs to read, because a request that ignores it will be refused rather than negotiated.

Two consequences worth stating plainly to anyone building against this. An unknown scope resolves to tier 4, the most restrictive — a typo fails closed, never open. And a write to any scope at tier 3 or above is refused outright today rather than stored, because the encrypted store those fields belong in is not yet wired. We would rather refuse than accept something we cannot yet protect.

Ask for a reason, not a list of fields.

A person cannot meaningfully consent to nineteen dot-paths. They can consent to “so the hotel can have your room the way you like it before you arrive.” A purpose is a named, pre-declared bundle: an audience, a plain-English ask, a TTL, and the exact scopes it resolves to.

The generator refuses to build a purpose that names an unknown scope, or that carries an intimate-tier scope at all. A bundle cannot quietly widen.

Composed over the rails you already speak.

PCHP does not replace MCP or A2A. MCP standardises how an agent reaches a tool; A2A standardises how two agents negotiate. PCHP adds the four things neither carries: who asked, under what consent, for what purpose, and with what receipt.

Following the multi round-trip pattern MCP adopted in its 2026-07-28 revision, an ungranted request is answered rather than blocked — the issuer names exactly what is missing and where to grant it, and the requester comes back. No held connection.

# 1. Read the vocabulary. Cache against its version.
GET /pchp/scopes.json          -> { version: "0.5.0", scopes: [...] }

# 2. Ask by purpose, not by field list.
POST /pchp/read
PCHP-Purpose: hospitality.stay
PCHP-Protocol-Version: pchp/2026-07-12

# 3. Not granted yet. You are told precisely what is needed.
{
  "resultType": "consent_required",
  "purpose": "hospitality.stay",
  "required": [ { "scope": "prefs.room.temperature", "tier": 2 } ],
  "grantAt": "https://hushh.ai/@ada/consent?req=01J...",
  "expiresIn": 900
}

# 4. The person decides. You retry with the grant.
POST /pchp/read
PCHP-Receipt: rcpt_01J...

# 5. You receive only the granted scopes, and a receipt exists on both sides.

Scope, purpose and receipt travel in headers deliberately: an enforcement point can allow, deny, meter and log an access without ever parsing the body. The gateway sees that a subscriber asked for a tier-2 scope under a named purpose, and never sees the person’s data.

Where this actually runs.

The order is not incidental: on device first, edge second, cloud last, and only when the work earns it. Apple Silicon shares one physical pool of memory across CPU, GPU and Neural Engine, which means the expensive operation in a hybrid personal-AI system is not compute — it is copying, and Apple already deleted the copy.

So intent parsing, redaction, rendering a consent card, and search over a person’s own vault belong on the phone: they must be instant and they must never leave. We are building toward the on-device stack Apple exposes — MLX, the Foundation Models framework, Core ML and the Neural Engine — with Private Cloud Compute’s guarantees as the reference posture for anything that does leave.

Day zero is Google-first and Gemini-first: someone arriving with a Google account and nothing else must have a working agent before configuring anything. Beyond that, every open model meeting the open-source community’s own standards for safety and trust is supported — a model you can inspect, host and keep is the only kind that fully honours own your AI.

What we are building toward, stated honestly.

For enterprise agents the integration targets are deliberate: Salesforceas the priority CRM surface, because that is where a human agent’s customer relationships already live, and MuleSoft as the API and integration layer beneath it. On the agent rails, MCP and A2A for tools and negotiation, AP2 and UCP for anything that moves value.