GlossaryMCP (Model Context Protocol)
MCP (Model Context Protocol)
The Model Context Protocol (MCP) is an open standard that defines how an AI application connects to external tools and data sources.
Anthropic released it as open source on 25 November 2024, then handed it on 9 December 2025 to the Agentic AI Foundation, a fund under the Linux Foundation, and it is now supported by assistants such as Claude or ChatGPT and by development tools such as Visual Studio Code. Its documentation compares it to a USB-C port: a software vendor writes the server that exposes its functions once, and any compatible application can plug into it, where previously a connector had to be written for every pair of tools.
How it works
A host application, the assistant or the agent, opens one connection per MCP server it plugs into, and each server offers three kinds of element. Tools are functions the model can call to act, such as creating a record, sending a file or querying a database. Resources are context data, such as the content of a document or a set of records. Prompts are reusable instruction templates. Messages follow the JSON-RPC 2.0 format, a local server communicates through the machine’s standard input and output, and a remote server goes through HTTP, with the documentation recommending OAuth to obtain access tokens. The current version of the specification is dated 28 July 2026.
What MCP solves and what it does not
The protocol standardises the plug, which is a real gain, since it moves the cost of integration from each pair of tools to each end. It says nothing, on the other hand, about the quality of what is exposed: a fully compliant server can offer read-only access, ignore custom fields or stop working when the software behind it changes its API. Plugging in creates no new right either, since the agent sees exactly what the account used for the connection sees. The useful question for a vendor is therefore not “do you speak MCP” but which objects it exposes for writing and with which rights, as our article on whether to require MCP and the one on what the protocol does not solve explain.
An MCP server is a supplier
The specification itself asks that tool descriptions be treated as untrusted unless they come from a trusted server, and that the user’s explicit consent be obtained before a tool is called. The reason is simple: a tool’s description is read by the model as an instruction, and an update to the server can change it without anything visible changing. A server also holds credentials to your systems, so it should be chosen like a supplier, and the company that runs a remote server handling personal data on your behalf is your processor under the GDPR. The three questions to ask before plugging one in are in our article on the MCP server as a supplier.
Balt connects to more than 3,200 tools, through MCP among other ways, and you decide what he can read and what he can write in each of them.
Sources
- Model Context Protocol, What is the Model Context Protocol (MCP)? (read 1 November 2026)modelcontextprotocol.io
- Model Context Protocol, Architecture overview, version 2026-07-28 (read 1 November 2026)modelcontextprotocol.io
- Model Context Protocol, Specification, version 2026-07-28, Security and Trust & Safety section (read 1 November 2026)modelcontextprotocol.io
- Anthropic, Introducing the Model Context Protocol (25 November 2024)anthropic.com
- Model Context Protocol, MCP joins the Agentic AI Foundation (9 December 2025)blog.modelcontextprotocol.io

