MCP adds authorization and a new transport
The March 2025 revision of the Model Context Protocol adds an OAuth 2.1 based authorization framework, replaces the old SSE transport and lets tools declare whether they only read or can change things.
The Model Context Protocol published a new specification revision on March 26, 2025. MCP is the open protocol that lets an AI assistant call tools and read data through a common interface. It is the first revision written with company deployments in mind.
What changed
- Authorization. The spec now defines how a client gets permission to use a server, built on OAuth 2.1. Before this, every server handled access its own way.
- Streamable HTTP. A new transport replaces the earlier HTTP plus server-sent events design, which makes remote servers simpler to host behind ordinary web infrastructure.
- Tool annotations. A tool can now say whether it only reads data or can change or delete it. A client can use that to decide when to ask a person before acting.
- Smaller additions. JSON-RPC batching and a completions capability for argument suggestions.
On the same day, OpenAI released version 0.0.7 of its Python Agents SDK with support for calling MCP servers. That is the first shipped MCP support from a second major model provider.
Why it matters for business data
An agent that can reach your systems needs two things before anyone should trust it with real work. It needs permission that is checked the same way every time, and it needs to know what a tool will do before it calls it. This revision starts on both.
What the protocol does not do is tell the agent what your data means. A tool can say it returns "customers". It cannot say which customers your finance team counts, or why sales counts differently. That part has to come from somewhere else, and without it the agent guesses.