MCP publishes the November 2025 specification

Background

MCP, the Model Context Protocol, is an open specification that describes how AI applications connect to external tools and data — a database, a code repository, a web service. Each revision of the specification is dated, so that people building clients and servers can be explicit about which version they are implementing and testing against.

The release

On 25 November 2025 the project published the November 2025 specification. The date was not incidental: the release came on the protocol's first anniversary^[1]. The document is a revision of a specification that had been in use for a year, so most of what it contains is about filling in gaps that implementers had run into rather than starting over.

What the revision adds

Task-based workflows. The release adds support for task-based workflows^[1]. Work that takes longer than a single request and reply — where a client needs to know how far along a job is, and a server needs somewhere to record progress — can now be described in the protocol itself instead of being improvised on top of it.

Authorisation, security and enterprise features. The revision simplifies authorisation flows and adds security and enterprise features^[1]. Authorisation is the part of a protocol that decides which client may reach which resource, and it is usually the most fiddly part to implement; simplifying the flow lowers the work needed to connect a server, while the added security and enterprise features target the settings where an organisation, not an individual, is doing the deploying.

A formal extension mechanism. The specification now defines a formal mechanism for extensions^[1]. Previously, anyone wanting behaviour beyond the core had to fork or patch around it; a defined extension path lets additions live alongside the standard without breaking clients that do not know about them.

URL mode elicitation. The release adds URL mode elicitation for secure out-of-band interactions with users^[1]. Elicitation is where a server asks the user for something it needs in order to continue; the URL mode moves that interaction out of the main channel, so the exchange happens in a separate, controlled place rather than inline.

Sampling with tools. The release adds sampling with tools, letting agentic servers call tools while sampling^[1]. Sampling is the mechanism by which a server asks the client's model to generate something; combining it with tools means a server that is driving a multi-step job can reach for a tool in the middle of that generation rather than pausing and resuming around it.

Adoption

The project also reported numbers on how far the protocol had spread in its first year. It said the MCP Registry held close to two thousand entries, a 407% growth over the first batch onboarded in September^[1]. It said active MCP servers had grown from a few experimental ones to thousands within the year^[1].

Why it matters

A specification's usefulness depends on both the standard and the population of things that implement it, and this release is addressed to both. The feature work — tasks, extensions, a simpler authorisation flow, safer ways to talk to the user — responds to the practical questions that appear once a protocol is used for real work rather than demos. The registry and server counts describe the other half: whether there is anything on the other end to talk to. One year in, the project is reporting both a revised document and an ecosystem measured in the thousands.