Skip to main content
Workflows Library MCP Directory Realtime AI News Sponsor Tier Subscribe
Front Page / AI News / Deep Dive

Hugging Face Launches Open-Agent Protocol 1.0: The Open-Source Standard for Agent Interoperability in 2026

Hugging Face released Open-Agent Protocol 1.0 today, an open-source agent communication standard backed by a Linux Foundation working group. OAP 1.0 combines MCP tool discovery, A2A agent-to-agent messaging, and a new Agent Card v2 format with capability negotiation.

Deepak Bagada

Deepak Bagada

CEO, SaaSNext

Aug 22, 2026 Published
|
Aug 22, 2026 Updated
|
6 Minutes Reading Time
Core Takeaways for Founders & Builders
  • OAP 1.0 merges MCP tool discovery, A2A agent messaging, and orchestration into a single protocol backed by Linux Foundation governance
  • Backward compatibility with MCP means existing 2,000 plus MCP servers work with OAP agents on day one through a translation proxy
  • The three-layer architecture addresses fragmentation but adds complexity for simple use cases where MCP alone suffices

Hugging Face released Open-Agent Protocol (OAP) 1.0 today, an open-source agent communication standard that merges three existing protocols into a single specification. The protocol combines MCP for tool discovery, A2A for agent-to-agent messaging, and Agent Cards for capability description. It is backed by a Linux Foundation working group with participation from Anthropic, Google, Microsoft, Meta, and over 40 AI infrastructure companies.

OAP 1.0 addresses the fragmentation problem in the agent ecosystem. Today, MCP handles tool-to-agent communication, A2A handles agent-to-agent messaging, and Agent Cards describe agent capabilities, but they are separate specifications with different transport layers, authentication models, and data formats. OAP unifies them into a single protocol stack.

Protocol Architecture

OAP 1.0 defines three layers:

Layer 1: Tool Discovery (MCP-compatible) Agents discover available tools through a standardized registry. The tool description format extends MCP tool schema with OAP-specific fields for capability negotiation, rate limits, and pricing. Backward compatibility with existing MCP servers is maintained through a translation proxy.

Layer 2: Agent-to-Agent Messaging (A2A-compatible) Agents communicate through a message bus that supports synchronous request-response and asynchronous pub-sub patterns. The A2A Agent Card format is extended with OAP Agent Card v2, which adds capability negotiation, trust level declarations, and cost estimation.

Layer 3: Orchestration (New) OAP 1.0 introduces a built-in orchestration layer for multi-agent workflows. Unlike LangGraph, which runs orchestration in application code, OAP orchestration runs in the protocol layer. Agents declare their workflow requirements in their Agent Card, and the protocol runtime handles scheduling, routing, and failure recovery.

Agent Card v2 Format

The Agent Card v2 includes oap_version (1.0), agent_id (uuid-v4), name, capabilities (tools array, models array, max_concurrent_tasks, supported_protocols), trust_level (verified), pricing (per_task cost in USD), authentication (oauth2 with issuer URL), and endpoint (OAP v1 URL).

Key Differences from MCP and A2A

MCP covers tool-to-agent scope with stdio/SSE transport, OAuth 2.1 auth, and built-in tool discovery. A2A covers agent-to-agent scope with HTTP/gRPC transport and built-in messaging. OAP 1.0 covers all three layers with HTTP/gRPC/WebSocket transport, OAuth 2.1 plus mTLS auth, built-in tool discovery and messaging, and built-in orchestration. OAP maintains backward compatibility with both MCP and A2A.

Linux Foundation Governance

The OAP specification is managed by the Linux Foundation Agentic AI Working Group. Founding members include Anthropic, Google, Microsoft, Meta, and Hugging Face. Contributing members include LangChain, CrewAI, AutoGen, and PydanticAI. Adopters include AWS, Oracle, IBM, Salesforce, and SAP. The working group follows a 6-month release cadence with OAP 1.1 expected in February 2027.

Adoption Path

For MCP server developers: OAP 1.0 maintains full backward compatibility. Existing MCP servers work with OAP agents through a translation proxy. No code changes required.

For agent framework developers: OAP orchestration layer is optional. LangGraph, CrewAI, and AutoGen continue to work as-is. OAP orchestration is an alternative for teams that want protocol-level workflow management.

For enterprises: OAP provides a vendor-neutral standard. Agents built against OAP work across AWS, Azure, GCP, and on-premise deployments without protocol translation.

Enterprise Impact

  1. Standard Convergence: OAP 1.0 reduces protocol fragmentation. Enterprises no longer need to choose between MCP and A2A.
  2. Vendor Portability: Linux Foundation governance prevents any single vendor from controlling the standard.
  3. Ecosystem Growth: OAP backward compatibility with MCP means existing 2,000 plus MCP servers work with OAP agents on day one.
  4. Adoption Risk: The three-layer protocol is more complex than MCP alone. Teams with simple tool-calling needs may find MCP sufficient.

Published: August 23, 2026. Protocol specification and working group details confirmed via Hugging Face blog and Linux Foundation announcement.

Executive Briefing

Enjoyed this breakdown? Get our morning dispatch in your inbox.

Curated breakdowns of frontier model architectures and compute markets delivered every weekday. Zero fluff.

🎉 Thank You for Subscribing!

Frequently Asked Questions
OAP 1.0 is a superset that includes both MCP and A2A functionality. Existing MCP servers and A2A agents work with OAP through backward-compatible translation proxies. Teams with simple tool-calling needs can continue using MCP directly. OAP adds value when you need agent-to-agent messaging and protocol-level orchestration.
MCP server developers can adopt OAP immediately with no code changes since it is backward compatible. Agent framework developers can optionally add OAP orchestration alongside existing LangGraph or CrewAI orchestration. Enterprise adoption typically follows a 6-12 month evaluation cycle with production deployments expected in Q2 2027.
Deepak Bagada
Author Profile

Deepak Bagada

CEO, SaaSNext

Deepak Bagada is the CEO of SaaSNext and founder of Daily AI World. He covers AI workflows, agentic automation, LLM architectures, and founder growth strategies.

Related Intelligence Analysis

Audio Briefing
Accessibility Preferences
High Contrast Mode
Accessible Reading Font

Keyboard Shortcuts

Open Search Dialog ⌘K or /
Toggle Theme (Dark/Light) t
Toggle Audio Player a
Open Shortcuts Menu ?
Close Active Dialog Esc