English 中文版
The Neural Protocol Suite is a family of five coordinated protocols that together replace the HTTP/REST stack for AI-native workloads:
| Layer | Name | Analogue |
|---|---|---|
| NCP | Neural Communication Protocol | Wire format / framing (think HTTP/2 frames) |
| NWP | Neural Web Protocol | Request / response (think HTTP semantics) |
| NIP | Neural Identity Protocol | TLS / PKI for agents |
| NDP | Neural Discovery Protocol | DNS for agents |
| NOP | Neural Orchestration Protocol | SMTP + MQ — multi-agent DAGs |
All five share a single default port — 17433 — and route on a 1-byte frame type. A single connection can carry frames from any layer.
HTTP and REST assume a human reads the response. For an AI agent, the assumptions are wrong in four specific ways:
A REST endpoint re-describes its schema implicitly in every response (field names, type hints, nesting). For an agent that queries 100 products, the same field names travel 100 times.
NPS’s AnchorFrame publishes the schema once. Agents reference it by a SHA-256 content-addressed id — sha256:abc…. The server can invalidate it or evolve it with a DiffFrame.
HTTP identity is a patchwork: cookies, OAuth, API keys, mTLS. For a multi-agent workflow where Agent A delegates to Agent B, there is no standard answer for “does A actually speak for the user?”
NPS’s IdentFrame and TrustFrame make Ed25519 identity first-class. Every frame can be signed; the NipCa issues and revokes identities with a spec’d certificate lifecycle.
REST APIs ship OpenAPI schemas that describe types, not meaning. An agent has to infer that price means “the retail price in USD” from context.
NPS schema fields carry semantic annotations — e.g. commerce.price.usd — drawn from a shared ontology. Agents can resolve intent without prose analysis.
REST is request-response. Streaming is a 2009 bolt-on (SSE / chunked transfer). Multi-agent orchestration is left entirely to frameworks (Temporal, Airflow, LangGraph).
NPS makes streams (StreamFrame, AlignStreamFrame) and orchestration (TaskFrame, DelegateFrame, SyncFrame) wire-level primitives.
| You are building… | NPS helps you by… |
|---|---|
| A knowledge base / API for AI shoppers | Publishing an anchor once → 80% token reduction per query |
| A multi-agent orchestrator | Using TaskFrame DAGs with signed delegation chains |
| An agent marketplace | Cryptographic identity + discovery without running your own DNS |
| A compliance-bound service | AaaS Profile L1 / L2 / L3 levels define what “production-ready” means |
| A tool that consumes agent-produced data | Schema-anchored responses instead of LLM-parsed JSON |
NPS is not a replacement for LLM APIs (OpenAI, Anthropic). It is the network those LLM-driven agents speak when they talk to each other, to data sources, and to each other’s tools.
a2a-bridge adapter is on the roadmap📖 For tutorials, references, and operator guides, see the NPS Wiki.