Every capability of the agentic communication protocol.

Four messaging patterns, React hooks, a C# node host, three auth modes, service discovery, and deep LLM provider support — all in one protocol.

Four Messaging Patterns

Choose the right communication pattern at the call site. ANCP routes and delivers each pattern correctly — no custom transport logic needed.

Fire-and-Forget

Returns 202 Accepted immediately. The caller is never blocked. Work executes in the background on the receiving node with no response channel needed.

  • Email notifications
  • Background job submission
  • Audit log writes
  • Event publishing

Request-Reply

Synchronous — blocks until a typed response is received. The protocol handles serialisation, error propagation, and correlation ID matching automatically.

  • Data queries
  • Calculations
  • Validation checks
  • Form auto-fill

Streaming (SSE)

Real-time chunk delivery via Server-Sent Events. The handler yields chunks as they are produced; the client receives them as they arrive.

  • LLM text generation (tokens live)
  • Report generation with progress
  • Large data exports
  • Chat applications

Task-Start

Returns a taskId immediately. The task runs asynchronously; the caller polls for state: Running, Completed, Failed, or Cancelled.

  • Batch processing (hours)
  • Large analysis jobs
  • Background workflows
  • Long-running AI operations

LLM Provider Support

One protocol connects to any language model. Per-model cost tracking — TextInputCost and TextOutputCost per 1M tokens — stored automatically for every call.

Anthropic Claude

Recommended provider for agentic workloads. Supported models:

  • claude-3-5-sonnet-20241022
  • claude-3-opus-20250219
  • claude-3-sonnet-20240229
  • claude-3-haiku-20240307
  • claude-2.1

OpenAI GPT

Full GPT model family with unified request normalisation. Supported models:

  • gpt-4o
  • gpt-4-turbo
  • gpt-4
  • gpt-3.5-turbo

Self-Hosted / Local

Run your own models with zero data leaving your infrastructure. Supported runtimes:

  • Ollama (llama2, mistral, neural-chat)
  • HuggingFace endpoints
  • Custom fine-tuned models
  • Any OpenAI-compatible endpoint

React Integration (ancp-react)

Install @bizfirst/ancp-react and get fully typed, reactive hooks for every messaging pattern. No boilerplate — the provider handles auth, connection, and state.

AncpProvider

Wraps your app with the ANCP React context. Configure once at the root — baseUrl, nodeId, and auth. All child hooks consume this context automatically.

useAncpInvoke

Typed hook for request-reply invocations. Returns invoke, isLoading, data, error, and reset. Generic over payload and result types for full TypeScript safety.

useAncpStream

Streaming hook for SSE-backed actions. Exposes start, stop, chunks, isStreaming, and isDone — purpose-built for live LLM output in chat UIs.

useAncpTask

Long-task hook that manages task lifecycle automatically. Provides start, cancel, status (state, progress), and isRunning — handles polling under the hood.

useAncpCapabilities

Queries the discovery endpoint and returns the list of all available actions and their patterns. Enables dynamic UIs that adapt to the available node capabilities at runtime.

C# Node Backend

The BizFirst.Ancp.Node host handles HTTP transport, envelope parsing, authentication, and routing. You register typed handlers — ANCP does everything else.

OnInvoke

Register a request-reply handler: Ancp.OnInvoke<TPayload, TResult>('action', async ctx => {...}). Returns a typed result synchronously to the caller.

OnFireAndForget

Register a one-way handler: Ancp.OnFireAndForget<TPayload>('action', async ctx => {...}). Caller receives 202 immediately; handler runs to completion asynchronously.

OnStream

Register a streaming handler: Ancp.OnStream<TPayload, TChunk>('action', async ctx => { yield return chunk; }). Chunks are pushed to the SSE stream as the handler yields them.

OnTask

Register a long-running task handler: Ancp.OnTask<TPayload, TResult>('action', async (ctx, ct) => {...}). Returns a taskId immediately; the task is tracked and pollable.

AncpActionContext

Every handler receives a fully typed context object with: Payload, NodeId, CallerNodeId, TenantId, CallerTenantId, MessageId, CorrelationId, Timestamp, and Extensions.

Security & Multi-Tenancy

Enterprise-grade security that does not require custom middleware. Every envelope carries the full identity and tenant context needed for isolation, tracing, and billing.

Three Auth Modes

Choose the authentication model that fits your deployment:

  • JWT — bearer token, standard OIDC flows
  • API Key — simple secret header, great for service-to-service
  • DID — Decentralised ID for autonomous agents and cross-org federation

Tenant Isolation

Every IEnvelope carries TenantId and CallerTenantId. The node host enforces isolation automatically — no tenant data ever crosses a boundary, even in shared infrastructure.

Payment & Audit

Built-in token cost tracking (TextInputCost + TextOutputCost per 1M tokens), payment verification for metered AI services, and a full immutable audit trail on every operation. HTTPS/TLS 1.3 required.

Service Discovery

Every ANCP node publishes a machine-readable discovery document. Clients — and other agents — can query the node to learn what it can do before calling it.

Discovery Endpoint

GET /.well-known/ncp.json returns the AncpDiscoveryDocument — a complete listing of all registered actions and their messaging patterns. Cached for performance (100ms typical).

System Actions

Every ANCP node exposes built-in system actions out of the box:

  • ping — health check
  • getCapabilities — list all actions
  • getStatus — node status and metrics
  • getPaymentInfo — billing and usage data
  • pay — initiate payment for metered service

Agent Framework

ANCP carries the full Octopus AgentComposite — system prompt, capabilities, tools, conversations, memory, workflows, hooks, metrics, and audit log.

6 Agent Types

Every ANCP node can host agents of the following types:

  • Conversational — chat-first, context-aware
  • Analytical — data reasoning and insight generation
  • Generative — content and artefact creation
  • Reasoning — multi-step logical planning
  • Autonomous — goal-directed, self-directed execution
  • Custom — fully user-defined behaviour

5 Agent Roles

Roles define how agents interact with each other in multi-agent systems:

  • Operator — executes actions and tasks
  • Analyst — interprets and summarises data
  • Manager — coordinates other agents
  • Monitor — observes and reports on system state
  • Reviewer — validates and approves outputs

Platform Integration

ANCP is not a standalone library — it is the communication backbone of the entire BizFirstAi platform.

Octopus

ANCP is the communication backbone of Octopus. The chat interface sends messages via ANCP, the agent executor receives via ANCP, and streams the response token-by-token back via ANCP SSE. Every step is correlated and audited.

ProcessServer

AI Agent Nodes and AI Function Nodes inside ProcessServer call LLMs via ANCP actions. The full request lifecycle — auth, multi-tenancy, cost tracking, audit — is handled by the ANCP transport, not custom node code.

FlowStudio

Visual workflow nodes in FlowStudio can invoke any ANCP action as a workflow step. Drag an ANCP action onto the canvas, map the payload fields, and the visual designer handles the rest.

Competitive Advantages

ANCP was designed to address the gaps in existing transport and framework choices for agentic systems.

vs gRPC

ANCP is browser-compatible out of the box — standard fetch and SSE, no special networking, no Protobuf toolchain. JSON serialisation keeps payloads human-readable and debuggable. Auth is simpler and built in.

vs Message Queues

No broker infrastructure to deploy and maintain (no RabbitMQ, no Kafka cluster). ANCP provides request-reply natively, real-time streaming, and built-in payment verification — none of which message queues offer.

vs Direct LLM APIs

Direct LLM APIs lock you to one provider. ANCP adds multi-provider routing, workflow integration via ProcessServer, multi-tenancy, on-premise deployment capability, and custom agent orchestration — without losing access to any model.

Roadmap

ANCP is evolving toward autonomous multi-agent networks and decentralised AI economies.

Phase 2 — Q2-Q3 2026

Next-generation autonomous capabilities coming mid-2026:

  • Tool composition (agents call tools returning tools)
  • Autonomous agent loops
  • Multi-agent coordination
  • Bidirectional streaming (WebSocket upgrade)
  • OpenTelemetry integration
  • RAG pipeline as ANCP actions

Phase 3 — Q4 2026

Decentralised AI agent marketplace and cross-org federation:

  • Federated cross-org node networks
  • Cryptocurrency payments (USDC, USDT)
  • Escrow and multi-sig payment flows
  • AI agent marketplace with capability discovery

Build on the protocol designed for agents.

Production ready. Enterprise grade. v1.0.2 available now.