MCP · A2A · ACP · ANP · AG-UI · LMOS · AGNTCY · AITP — how agents talk to tools, each other, users, and the network. The full landscape as of Q1 2026.
These protocols are not competing — they address different layers, analogous to how HTTP, WebSocket, and gRPC coexist in web infra. The industry is converging under Linux Foundation governance. ACP already merged into A2A (Aug 2025). The winning architecture combines MCP for tool access, A2A for agent collaboration, ANP/AGNTCY for discovery, and AG-UI for user interaction.
| MCPAnthropic | A2AGoogle → LF | ANPCommunity | AG-UICopilotKit | LMOSEclipse Fdn | AGNTCYCisco → LF | AITPNEAR Protocol | |
|---|---|---|---|---|---|---|---|
| Identity & Core | |||||||
| Primary Purpose | Agent ↔ Tool/Data standardization | Agent ↔ Agent collaboration & task delegation | Decentralized agent discovery, identity, secure networking | Agent ↔ Frontend UI bi-directional interaction | Full-stack agent runtime & orchestration platform | Internet of Agents infra stack (discovery, routing, composition) | Agent interaction with payments & economic transactions |
| Origin / Backer | Anthropic (Nov 2024) | Google Cloud (Apr 2025) → Linux Foundation (Jun 2025) | ANP-Community; W3C White Paper (May 2025) | CopilotKit (May 2025) | Eclipse Foundation (Mar 2025) | Cisco Outshift + LangChain (Mar 2025) → Linux Foundation (Jul 2025) | NEAR Protocol (Feb 2025) |
| Status (Q1 2026) | De facto standard | Leading standard 50+ partners | Maturing nearing finalization | Production 10% of F500 | Active dev | Active dev | Early / niche |
| Governance | Anthropic (open spec) | Linux Foundation AI & Data | Community-driven (open source) | CopilotKit (open source) | Eclipse Foundation | Linux Foundation | NEAR Foundation (MIT license) |
| Industry Partners | Broad: OpenAI, Google, MSFT, etc. all support MCP clients | Google, AWS, MSFT, Salesforce, SAP, ServiceNow, IBM, MongoDB, 50+ | Community contributors; academic interest growing | MSFT Agent Framework, Google ADK, AWS Strands, LangGraph, CrewAI, Oracle | Eclipse ecosystem, Deutsche Telekom | Cisco, LangChain, Galileo | NEAR ecosystem primarily |
| Architecture & Transport | |||||||
| Wire Protocol | JSON-RPC 2.0 over stdio / SSE / HTTP |
JSON-RPC over HTTPS; SSE for streaming |
HTTP + JSON-LD; E2E encrypted channels |
JSON event stream over HTTP POST + SSE; optional binary |
WebSocket subprotocol + HTTP; pub/sub message bus |
gRPC + HTTP/2 (SLIM messaging) |
JSON-RPC over HTTP; designed for on-chain bridges |
| Data Format | JSON (structured tool schemas) | JSON; Parts with MIME types (text, files, data) |
JSON-LD (semantic linked data) |
Typed JSON events (16 event types) | JSON-LD (W3C Thing Descriptions) |
JSON-LD; semantic descriptions | JSON; NEAR-native data structures |
| Topology | Client-Server (1 host ↔ N MCP servers) | Client-Server (client agent → remote agent) | Peer-to-Peer (fully decentralized) | Client-Server (frontend ↔ agent backend) | Star topology (central Router + Registry) | Decentralized mesh (agent discovery service) | P2P with optional on-chain settlement |
| Streaming Support | Basic message streaming | SSE for real-time task updates; push notifications | Meta-protocol negotiation can select streaming transport | Native — built entirely around streaming events; token-level granularity | WebSocket subprotocols for real-time | gRPC streaming | Basic polling / webhook |
| Discovery, Identity & Security | |||||||
| Agent Discovery | MCP servers expose tool manifests; clients query them | Agent Cards (JSON) — name, skills, auth, endpoint URL | Decentralized discovery via DID documents + capability ads | N/A (frontend discovers agent via config endpoint) | Central Agent Registry + mDNS for local nets | Agent Discovery Service (ADS) — decentralized | Agent profiles registered on-chain or via API |
| Identity Model | Server identity via transport config | Agent Cards with OAuth 2.0 / API keys | W3C DID — decentralized identifiers; verifiable credentials | Session-based auth tokens | W3C DID + OAuth 2.0 (dual support) | W3C DID + VC (verifiable credentials) | NEAR account identity (crypto-native) |
| Auth / Security | Transport-level (process isolation, OAuth for remote) | OAuth 2.0; capability scoping; HTTPS required | E2E encryption; DID-based mutual auth; post-quantum ready (roadmap) | CORS, auth tokens, audit logs; guardrails at boundary | Encrypted comms; multi-scheme auth | Capability tokens; DID/VC-based; zero-trust | Crypto-native signing; on-chain verification |
| Task Management & Workflow | |||||||
| Task Lifecycle | Request → Response (stateless tool calls) | Full lifecycle: submitted → working → input-required → completed/failed/canceled | Not defined at protocol level (application layer concern) | RUN_STARTED → events → RUN_FINISHED (lifecycle signals) | INITIALIZING → ACTIVE → DEGRADED → RETIRING → RETIRED | Task routing + composition pipelines | Request → negotiation → settlement |
| Multi-turn / Long-running | Session context via MCP server state | Yes — tasks persist; agents can request human input mid-task | Meta-protocol can negotiate session parameters | Yes — shared mutable state; thread/run IDs; cancellation | Yes — stateful sessions survive restarts | Yes — workflow composition | Multi-step negotiation flows |
| Task Delegation | Host routes to appropriate MCP server | Core feature: client agent delegates to remote agent; hierarchical | P2P — any agent can request from any other | N/A (user↔single agent scope) | Scheduler/Router handles delegation | Pipeline composition; agent chaining | Contract-based delegation |
| Human-in-the-Loop & UX | |||||||
| HITL Support | Not in protocol spec | input-required state pauses task for human input |
Not in spec (application concern) | First-class: approval flows, tool confirmation, state sync, generative UI | Not specified | Not specified | Transaction approval inherent |
| UI / Frontend | N/A (backend protocol) | UX negotiation (modality-aware: text, audio, video, forms) | N/A | Core purpose: React hooks (useAgent), CopilotKit components, A2UI support, generative UI specs |
Multi-channel (web, mobile, voice IVR) | Not focused on UI | Wallet-based UX |
| Ecosystem & Interop | |||||||
| Framework Support | LangChain, LlamaIndex, CrewAI, Mastra, Cursor, Claude, GPT — ubiquitous | LangGraph, CrewAI, AG2, BeeAI, Vertex AI, most major frameworks | Reference implementations; growing | LangGraph, CrewAI, Mastra, Pydantic AI, Agno, LlamaIndex, MSFT Agent Framework, Google ADK | LangChain, LlamaIndex via LMOS adapters | LangChain native; Galileo eval | NEAR-specific SDKs |
| MCP Compatibility | — | Complementary: MCP for tools, A2A for agent collab | Can layer on top of MCP for tool access | CopilotKit supports MCP Apps natively | Can integrate MCP servers via protocol bindings | Can compose with MCP tools | Independent |
| SDK Languages | Python, TypeScript, Java, C#, Go, Rust, Kotlin | Python, TypeScript, Java, Go (via LF ecosystem) | Python (reference) | TypeScript, Python, Kotlin, Go, Java, Rust, .NET | Kotlin (JVM-native), Python | Python primarily | TypeScript, Rust (NEAR-native) |
| Edge / IoT Fit | Moderate — JSON-RPC overhead | Moderate — JSON-RPC over HTTPS | Good — P2P, lightweight identity; but JSON-LD parsing cost | Not designed for — frontend-oriented | Good — multi-transport (MQTT, CoAP supported) | Good — gRPC efficient; but cloud-centric design | Not designed for |
| Strengths & Weaknesses | |||||||
| Killer Feature | Universal tool integration standard; massive ecosystem adoption | Agent Cards for discovery; enterprise-grade task lifecycle; 50+ partners from day 1 | True decentralized P2P; DID-native identity; meta-protocol negotiation | Only protocol purpose-built for agent↔UI; streaming state sync; guardrails at boundary | K8s-native; full lifecycle mgmt; multi-transport flexibility | Composable agent pipelines; gRPC efficiency; eval-integrated (Galileo) | Native economic transactions between agents; on-chain settlement |
| Biggest Weakness | Not for agent↔agent; no task lifecycle; no discovery standard | Cloud/enterprise-centric; no edge optimization; JSON-RPC overhead | Early ecosystem; JSON-LD parsing expensive on constrained devices | Narrow scope (only UI layer); vendor-specific (CopilotKit) | Complex K8s dependency; smaller ecosystem than A2A | Cloud-centric; not addressed edge/IoT; smaller community | NEAR-ecosystem locked; limited adoption outside crypto |
| Maturity Score | ★★★★★ | ★★★★☆ | ★★★☆☆ | ★★★★☆ | ★★★☆☆ | ★★☆☆☆ | ★★☆☆☆ |
The USB-C of agent infra. If your agent needs to connect to databases, APIs, file systems, or any external tool — MCP is the standard. Already won this layer.
The winner for agent↔agent comms. With ACP merged in, Linux Foundation governance, and 50+ partners — this is the standard to bet on for multi-agent systems. Agent Cards are the discovery primitive.
The only protocol designed for a truly decentralized agentic web. DID-native identity, P2P discovery, meta-protocol negotiation. Ambitious but early. Watch this space for open agent marketplaces.
If you're building user-facing agentic products, this is your protocol. 16 event types cover everything: streaming, state sync, HITL, generative UI. CopilotKit + AG-UI + A2UI is the frontend agent stack.
Full platform play from Eclipse. Best for enterprises wanting K8s-native agent orchestration with multi-transport support (HTTP, MQTT, CoAP). Heavier lift but production-grade lifecycle management.
Cisco's composable agent infra stack. gRPC-based SLIM messaging for efficiency. Integrated with Galileo for eval. Still cloud-centric; strong on pipeline composition.
Niche but uniquely positioned for agent-to-agent payments and economic interactions. Built on NEAR. Only relevant if you're building agents that transact value on-chain.
For most agentic systems, combine these four protocols to cover every layer:
This mirrors how modern web infra layers HTTP + WebSocket + gRPC + DNS. Each protocol owns its layer; none tries to do everything.