Standard · Converse
MCP (Model Context Protocol)
How an agent calls your tools, reads your resources and uses your prompts.
What it is
MCP is a common way for an agent to use outside capabilities. A server offers tools (things an agent can do), resources (things it can read) and prompts. Any MCP client can then use any MCP server. A company can run one to let agents ask about it or act with it.
An analogy
At a glance
- Owner
- Created by Anthropic; given to the Linux Foundation’s Agentic AI Foundation in December 2025
- Maturity
- Emerging The 2026-07-28 revision is final. It is stateless (no initialize step, no session id) and uses server/discover instead. It formalizes extensions and hardens authorization. The MCP Registry, the public list of servers, is in preview.
- Where it lives
- modelcontextprotocol.io
- Framework domain
- Converse: Be asked
- What it serves
- Answer agents’ questions with live data
- Offer actions agents can call
Who should care
The roles this standard touches most. See all the roles on the personas page.
- Engineering and product. Build the access, data, endpoints and actions that agents use, on open standards.
- Service and operations. Answer agents’ questions, hand off to people well and keep promises made through agents.
- Commerce and sales. Make offers, prices, availability and checkout work for agents buying on someone’s behalf.
How to adopt it
- 1
Start read-only: tools for your facts, products or opening hours.
- 2
Add action tools such as apply, book or request a quote, each with a clear description.
- 3
Support the current revision and the one before it while clients move.
- 4
Publish a server card and list the server in the MCP Registry under a namespace you can prove.
The Institute’s view
The most widely supported route to being callable. The protocol is still changing: the stateless revision is recent, and the registry is a preview. Build small, keep the tools few and well described, and expect to update.
Standards change fast. Last reviewed November 2026.