Anthropic releases MCP 2.0 — multi-agent connection protocol gets standardized tool-call schema and cross-runtime auth handoff
Anthropic shipped MCP 2.0 this week, the second-generation Model Context Protocol with cross-runtime auth handoff and a standardized tool-call schema. The 2.0 release lets an agent in Claude hand off mid-task to an agent running in GPT-5 or Gemini 2.5 without re-authenticating against the connected tool surface. The protocol is the closest the agent ecosystem has come to a TCP/IP-style interoperability standard.
The cross-runtime handoff is the headline feature. MCP 1.x already let a single agent runtime call into shared tool surfaces (Slack, GitHub, Notion, internal databases). MCP 2.0 adds the auth-token-handoff layer that lets that same tool session migrate between agent runtimes — meaning a multi-agent pipeline can route a sub-task to whichever model handles that step best, then return to the orchestrator without losing tool state.
The standardized tool-call schema is the more durable contribution. Until now, every model lab maintained its own JSON tool-call format and required adapter shims for cross-vendor work. MCP 2.0 freezes a schema that GPT, Claude, and Gemini agents all emit and parse natively, which makes multi-agent orchestration tooling tractable for the first time outside single-vendor walled gardens.
Anthropic — MCP 2.0 release notes → · MCP.org — Protocol 2.0 specification → · Wired — The agent protocol everyone agreed on →