AI・テクノロジー
Anthropic and OpenAI Jointly Publish MCP v2.0 Specification — Standardizing Agent-to-Agent Communication Drives Down Integration Costs
機械翻訳 / Machine-translated
On August 18, 2026, Anthropic and OpenAI jointly published a draft specification for Model Context Protocol (MCP) version 2.0 on GitHub. The core features are a unified format for tool-call definitions and support for asynchronous streaming. If AI agents from different vendors can interoperate directly over a common protocol, this could mark a structural turning point that significantly reduces the "integration cost" of multi-agent architectures.
At 17:00 UTC on August 18, 2026, a draft PR for MCP v2.0 was merged into Anthropic's GitHub repository modelcontextprotocol/specification. Three OpenAI engineers were listed as co-reviewers on the PR, making this effectively a jointly authored specification.
The three main changes are as follows:
"With MCP v2.0, you no longer need an extra wrapper to plug Claude and GPT into the same pipeline. The era of writing throwaway integration code is over." (CTO of a SaaS startup, via X)
MCP v1.0 was published by Anthropic alone in November 2024. Originally designed as a tool-integration specification for Claude, it gained rapid adoption throughout the first half of 2025 as major code editors — including VS Code, Cursor, and Zed — announced compatibility. By the end of 2025, the number of registered MCP-compatible servers had surpassed 1,200, and the protocol had begun functioning as a de facto industry standard.
However, v1.x had a known incompatibility with OpenAI's tool-calling specification (Function Calling / Responses API). In cross-vendor multi-agent setups, teams frequently had to build their own "MCP ↔ OpenAI API translation layer," which was repeatedly cited as a source of engineering friction.
The v2.0 draft is positioned as an effort to resolve that "translation cost" at the specification level.
At the time of the draft's publication, only Anthropic and OpenAI are named, and no engineers from Google DeepMind were involved in the PR. Gemini has its own "Function Declarations" specification, and no timeline for conformance with v2.0 has been announced. The scenario in which the industry splits between MCP v2.0 adopters and those following Google's proprietary specification remains on the table.
Both Azure AI Foundry and Amazon Bedrock completed their MCP v1.x support within 2025. Their v2.0 roadmaps are expected to be announced within Q3 2026, and the pace of adoption by major cloud providers will largely determine how quickly v2.0 spreads.
A Gartner survey conducted in June 2026 (targeting Fortune 500 companies) found that 62% of respondents cited "protocol incompatibility" as the biggest barrier to adopting multi-vendor AI agents. If v2.0 achieves widespread adoption, it is expected to remove much of this barrier at the specification level.
The speed at which MCP v1.0 went from "Anthropic's standard" to "the industry's de facto standard" exceeded expectations. Looking at the curve of registered server counts going from zero to over 1,200 in 18 months, it seems likely that the mere fact that v2.0 was "co-authored by two competing companies" will prompt some organizations to accelerate their adoption decisions.
One important caveat: this is still a draft. The target for official RFC status is Q4 2026, and the possibility of breaking changes cannot be ruled out. It would be prudent to wait until the final release before designing production systems around v2.0.
For startups and in-house engineers, however, reading the v2.0 draft now could meaningfully shift architecture decisions six months down the road. The cost of reading it has never been lower than it is right now, before the standard is finalized.
The next things to watch are Google's response and the disclosure of AWS Bedrock's v2.0 roadmap. Those two developments will determine the true pace of MCP v2.0 adoption.
The joint publication of the MCP v2.0 draft could mark the moment a "common language" for AI agent communication begins to take shape. However, Google's absence and the draft's preliminary status call for caution before making immediate design changes. If your agent pipeline is already running on MCP v1.x, it's worth taking time this week to understand what Async Streaming and Context Propagation in v2.0 will actually change for you.
How does your system connect agents from different vendors?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
