Ask for a capability
The app describes the job, such as answering a question, creating an image, or processing speech.
What is an AI mesh?
An AI mesh is a way for AI programs to find the right helper without knowing its address in advance.
An AI mesh is a network layer that helps autonomous agents discover available AI capabilities, select a suitable provider, and deliver work without hardcoding every provider endpoint.
Definition
An AI mesh is a shared network for AI helpers. An app says what it needs, finds a suitable available provider, and sends the task directly to it.
Definition
An AI mesh connects agents, models, tools, and compute providers through shared discovery and routing rules. Instead of assuming one vendor or known endpoint, a client asks which healthy provider can satisfy an intent and then communicates with the selected provider directly.
From agent intent to provider response
IICP separates the control plane used for discovery from the data plane used for task execution. This keeps the directory out of the task payload path.
- 1Express an intentThe client identifies the required capability and optional constraints.
- 2Discover providersThe directory finds public-routable nodes advertising a compatible intent.
- 3Select a routeThe client evaluates the returned providers and routing signals.
- 4Send the task directlyThe selected provider receives and executes the task.
Ask, choose, connect.
Choose an available helper
The network finds providers that say they can perform the job and are currently available.
Connect directly
The app sends the task to the chosen provider. The directory does not need to carry the task itself.
Why support it? It helps AI software move away from one hardcoded provider and toward an open network where different helpers can be found, compared, and used when needed.
What an AI mesh needs
Capability registration
Providers publish the tasks, models, limits, regions, and transports they can support instead of relying on a fixed application-specific address.
Agent discovery
A client describes an intent and constraints. The discovery layer returns providers that currently advertise a matching capability.
Health and liveness
Heartbeats and reachability observations help prevent stale or unavailable providers from being treated as healthy routing targets.
Provider selection
Matching providers can be ordered using operational signals such as health, load, reputation, region, latency, and policy constraints.
Direct task delivery
Discovery and execution are separate. After discovery, task traffic goes to the selected provider rather than through the IICP directory.
Protocol interoperability
An AI mesh should complement tool and agent protocols. IICP can route work beneath MCP, A2A, ACP, and OpenAI-compatible client surfaces.
AI mesh compared with adjacent systems
| System | Primary concern | Typical addressing | Relationship to IICP |
|---|---|---|---|
| Service mesh | Traffic policy between known application services | Service identity or hostname | IICP adds intent and capability discovery for AI providers. |
| Multi-agent framework | Workflow and agent orchestration | Configured agents and roles | A framework can use IICP to locate an external capable provider. |
| Model router | Selecting among configured models or vendors | Known model catalogue | IICP discovers independently operated nodes by advertised intent. |
| IICP AI mesh | Capability discovery and routing | Intent URN plus constraints | Open protocol and reference network in active beta. |
What exists today
Live
Registration and discovery
Provider registration, heartbeat, public registry views, intent discovery, and direct provider endpoints are implemented.
Beta
Routing and cooperative inference
Health, scoring, reputation, credits, proxy surfaces, and cooperative inference are active but can still change.
Planned
Federated control plane
The live network currently depends on the genesis directory. Open replica federation is a Phase 6 objective.
Current limitation: IICP is an open protocol with a live beta reference network, but the production discovery control plane is not yet fully federated. Review the known limitations before treating it as production-critical infrastructure.
Ask the mesh for a chat-capable provider
A client can query the discovery endpoint with an intent URN. The response supplies routable matching providers; the subsequent task is sent directly to the chosen node.
curl -sf \
'https://iicp.network/api/v1/discover?view=public&intent=urn:iicp:intent:llm:chat:v1'Discovery API reference →Read the protocol specificationBrowse mesh nodes
AI mesh FAQ
Is an AI mesh the same as a service mesh?
No. A service mesh manages traffic between known services. An AI mesh additionally needs capability descriptions, intent discovery, provider selection, and AI-specific execution contracts.
Does IICP carry every task through a central gateway?
No. The directory participates in discovery. After a provider is selected, the client sends the task directly to that provider endpoint.
Is the IICP mesh decentralized today?
The providers are independently operated, but the live discovery control plane currently uses one genesis directory. Federated directory replicas are planned and partially implemented, not yet the production default.