Scope, Goals & Non-Goals
What this specification covers, the retail dealership data and tool layer for AI agents, and the adjacent territory it deliberately leaves to others.
1.1 Purpose
This specification defines how retail automotive dealership systems expose their data and actions to AI agents over the Model Context Protocol (MCP). It establishes a common set of resources, tools, naming conventions, and a security profile so that any conformant agent can work with any conformant server without bespoke, per-vendor integration.
The protocol layer is MCP. This specification is the automotive retail profile of MCP: the agreed-upon resource and tool surface for the dealership.
1.2 Goals
- Make any conformant dealership server interchangeable from an agent's point of view.
- Cover the operational core of the dealership: inventory, leads and CRM, deals and desking, service, parts, F&I, and marketing/analytics signals.
- Keep the surface small, stable, and versioned, so vendors can implement it once and rely on it.
- Treat security and PII as first-class, automotive data carries real compliance weight.
- Be implementable on top of existing DMS, CRM, and platform APIs without forcing a re-platform.
1.3 Non-Goals
This specification is the retail / dealership data-and-tool layer. It is explicitly distinct from, and does not attempt to replace:
- Vehicle and OEM signal standards such as COVESA / VSS (Vehicle Signal Specification). Those describe signals from the vehicle itself; this describes the systems of the dealership.
- The Model Context Protocol itself. MCP is the transport and capability model; this profile rides on top of it and does not redefine it.
- Proprietary vendor APIs. A conformant server may be a thin adapter in front of an existing DMS or CRM API. This spec standardizes the agent-facing surface, not the system of record.
Claiming the dealership, not the vehicle, is deliberate: it avoids confusion with vehicle-signal efforts and gives implementers an unambiguous target.
Related work: WebMCP
WebMCP is a second, complementary standard, and it is not in tension with this one. It answers
a different question: how a webpage exposes tools to an AI agent that is already inside the
browser (via document.modelContext), as opposed to how a backend dealership system exposes
tools to any agent over the network, which is what this profile and base MCP both describe. A
dealer's website is a webpage; WebMCP is how that page could offer "check availability" or "book a
test drive" directly to a browser-resident shopping agent, independent of whether the dealer's
DMS/CRM has a conformant server at all. A fully agent-ready dealer implements both surfaces.
This specification is aware of WebMCP but does not adopt or profile it, deliberately. WebMCP is early: it is W3C Web Machine Learning Community Group incubation, explicitly "not a W3C Standard nor on the W3C Standards Track," shipping only behind Chrome and Edge origin trials, with Firefox and Safari standards-positions still open. Standardizing its tool shapes is out of our lane and premature. We note where it fits and leave it there.
1.4 Audience
DMS, CRM, and website/platform vendors implementing a conformant server; AI-agent vendors consuming it; and the dealers and agencies who will ask "do you support MCP, and which version?"
1.5 Base protocol currency
This profile targets the 2026-07-28 revision of the Model Context Protocol for its transport
and authorization mechanics. That revision is a large one: it replaces the initialize /
initialized handshake and Mcp-Session-Id with a stateless core, adds a mandatory
server/discover RPC that carries a protocolVersion per request, adds cacheable list results
(ttlMs / cacheScope), and formalizes an Extensions framework under which a vertical like
this one can publish net-new capabilities under its own reverse-DNS namespace. See
Conformance §2.5 for how this profile uses that framework, and
Reference §7.1 for a manifest written against the current discovery model.
MCP itself is now governed as "Model Context Protocol, a Series of LF Projects, LLC": on 2025-12-09 it was donated into the Agentic AI Foundation under the Linux Foundation. It is no longer an Anthropic-owned mark, and the naming constraints on "MCP" follow the LF Projects trademark policy rather than any single company's rule.
This section is a draft (RFC). Boundaries, especially the line between this profile and vehicle-signal standards, are exactly the kind of thing the council expects to refine with its partners.