/glossary

IICP glossary

These definitions keep public documentation, machine-readable resources, and implementation guidance consistent. The normative specification wins if a glossary summary and a protocol requirement ever disagree.

IICP provides discovery, eligibility, route authorization, and execution selection. It does not override a runtime's sandbox, network policy, tool permissions, authentication, consent, or operator access controls.

IICP: The Intent-based Inter-agent Communication Protocol: a protocol-neutral intent-resolution and execution-selection control plane. It discovers eligible providers and authorizes routes; it does not define every task protocol or place task payloads in the directory.
Intent: A stable identifier for the capability a consumer needs, usually expressed as an IICP URN such as urn:iicp:intent:llm:chat:v1. An intent describes purpose, not a specific provider address.
Consumer: A client or agent that requests a capability. The consumer applies its own policy and sends the task to the selected provider rather than to the directory.
Provider: A node or service that advertises capabilities and accepts authorized tasks at its execution endpoint. A capability claim does not by itself prove trust, health, quality, or permission.
Directory: The IICP control-plane service for registration, heartbeat, discovery, eligibility evidence, selection metadata, and route authorization. It is not an inference backend and is not intended to receive task payloads.
Capability: A declared function a provider can perform, described by an intent and related limits or compatibility metadata. Current availability and policy eligibility are evaluated separately.
Discovery: The process of finding candidate providers from advertised capabilities and constraints. Discovery does not automatically establish authorization or guarantee that an endpoint will succeed.
Intent resolution: The process of interpreting an intent and constraints to identify eligible execution candidates. It is not a general-purpose agent planner.
Provider selection: Choosing among eligible candidates using policy, compatibility, availability, operational evidence, and an explicit selection profile. Selection remains distinct from execution.
Dispatch authorization: Short-lived authorization binding an approved route to relevant parties, target, intent, and expiry. It does not grant broader access to the provider host, tools, network, or user data.
Execution protocol: The task contract used after selection, such as native IICP, an OpenAI-compatible endpoint, MCP, or a separately defined A2A binding. IICP resolution does not replace those application protocols.
IICP-CX: The optional payload-confidentiality profile. It can protect a request on its route to a key-ready provider, but the selected remote executor can still read plaintext while executing it.
Conformance evidence: A test result classified by who ran and verified it. Self-attested, project-verified, and independent evidence are different classes and must not be presented as equivalent.
AI mesh: An introductory description for a network of consumers and providers connected through discovery and routing. The phrase alone does not prove decentralization, diversity, availability, or independent operation.