Helpers say what they can do
Provider nodes join the network and list the kinds of jobs they support.
AI agent discovery lets one AI program find another AI helper that can perform the job it needs.
IICP helps a client find AI providers without knowing their addresses in advance. Providers advertise capabilities; clients discover candidates using a shared intent vocabulary.
AI agent discoverymeans finding the right AI helper for a task. The app asks what it needs, and IICP returns providers that advertise the capability and currently satisfy the directory's eligibility filters. Reachability is checked separately.
AI agent discovery is the process of finding an agent or provider that can perform a requested task under current constraints. IICP discovers provider nodes by intent URN, capability profile, operational state, and routing signals rather than requiring a fixed endpoint in application code.
Provider nodes join the network and list the kinds of jobs they support.
An app describes the capability it needs instead of naming one fixed provider.
The app receives suitable providers and decides which one to contact.
| Concept | Purpose | IICP surface |
|---|---|---|
| Intent registry | Defines the shared vocabulary and meaning of capability identifiers. | Intent Registry |
| Node directory | Maintains current provider registrations, health, public profiles, and routing data. | Node Directory |
| Discovery protocol | Defines how clients query for matching providers and receive routable candidates. | IICP specification |
The provider must advertise the requested intent URN or a protocol-defined compatible capability.
Expired, evicted, or unhealthy registrations should not be returned as normal routing candidates.
Clients can prefer compatible network exposure, geographic region, and supported task transport.
Capability profiles can describe model families, token limits, inference engines, and availability windows.
Operational observations can help order matching candidates, but clients should retain their own policy thresholds.
Credits, pricing declarations, conformance, and application policy can further constrain provider choice.
curl -sf \
'https://iicp.network/api/v1/discover?view=public&intent=urn:iicp:intent:llm:chat:v1'For public, privacy-reduced directory browsing, use /api/v1/registry/nodes. For task routing, use the discovery endpoint or an IICP SDK because discovery can return routable provider details.
PHP directory 1.10.93 and Rust directory 0.1.15expose optional, content-free explanations in discovery responses. Protocol suite 1.10.4 introduced the separation of a provider runtime's own name and version from its SDK compatibility version; the older sdk_versionfield remains a migration alias. SDKs from the 0.7.104 compatibility baseline expose evidence fields when the selected directory provides them. Older clients can ignore additive fields.
health_reasons separates reachability, backend, trust, and policy evidence.latency_evidence states whether an estimate comes from the multi-proxy EMA or has no accepted basis.sdk_release distinguishes protocol compatibility from relation to the latest version known by that directory release.trust_progress separates the Gold task threshold from other unmet Gold requirements.diversity_evidence reports only aggregates; it does not expose operator identifiers or claim independent failure domains.These explanations do not change eligibility, ranking, route exposure, or the client's responsibility to apply its own routing policy.
MCP and A2A define execution and interaction boundaries for tools and agents. IICP resolves an intent and selects an eligible provider using capability, policy, health, and network evidence. An orchestrator can then open the protocol-specific interaction it needs without sending the task payload through the directory.
Known-endpoint APIs such as OpenAI-compatible HTTP remain useful execution surfaces. IICP adds a discovery step before the endpoint is known.
Explore protocol interoperability →Return an explicit no-match result, relax only permitted constraints, or fall back to a configured local provider.
Apply heartbeat expiry and reachability checks so inactive nodes do not remain valid indefinitely.
Use bounded timeouts, circuit breaking, and the next eligible candidate rather than retrying forever.
Treat registration as a claim. Use conformance, identity, reputation, and task-level validation where risk requires it.
No. The intent registry can describe chat, embeddings, image, speech, tool, and future task categories. Availability depends on what active providers advertise.
No. It returns eligible candidates and routing signals. “Best” depends on client priorities such as latency, cost, quality, location, trust, and privacy.
IICP documents integration patterns for routing through an MCP-facing gateway and for exposing MCP-backed capabilities through an IICP provider.