In November 2024, two developers at Anthropic quietly released a protocol specification. No major conferences, no bright stages. The goal was to solve a problem that anyone who has integrated AI with external systems immediately recognizes: every new tool required a new connector, every model required a different integration. The result was throwaway code stacked on top of more throwaway code.
What those two developers didn't know at the time is that in less than eighteen months, Microsoft, OpenAI, Google and virtually the entire corporate AI ecosystem would adopt that protocol as a standard. Including Microsoft 365 Copilot.
The Model Context Protocol has arrived. And if you work with M365, it's worth understanding what changes.
The Nightmare Everyone Pretended Was Normal
For each combination of model and data source, a specific connector had to be written. Ten models, a hundred data sources: mathematically that was up to a thousand independent integrations to build and maintain. In the AI ecosystem, this became known as the "M×N problem": M models times N sources. It grew at the same pace as technical debt.
MCP solves this with a hub-and-spoke approach: you build an MCP server once, and any compatible client can connect. The analogy that stuck: USB-C for AI. Before, every device needed its own cable. After, a single standard serves everything.
The Protocol Born From a Developer's Frustration
On November 25, 2024, developers David Soria Parra and Justin Spahr-Summers launched MCP. The origin is wonderfully mundane: the protocol was born from Soria Parra's frustration with the repetitive task of manually copying code between Claude Desktop and his code editor. Small problem, big solution.
In the first months, adoption grew organically. By February 2025, the community had already created more than a thousand public MCP servers. In March 2025, OpenAI adopted the protocol in its Agents SDK. Sam Altman was direct: "People love MCP and we are excited to add support across our products." Competitors adopting a competitor's standard is the clearest possible signal that something has become a de facto standard.
In December 2025, Anthropic donated the protocol to the Agentic AI Foundation, created under the Linux Foundation, with co-founders including Anthropic, Block and OpenAI, plus support from Google, Microsoft and AWS. MCP stopped being "Anthropic's protocol" and became neutral ecosystem infrastructure.
Microsoft Showed Up to the Party. And Brought the Main Course.
Microsoft announced the general availability of MCP in Copilot Studio in May 2025, and expanded MCP tool call support to Microsoft 365 Copilot Chat, the Researcher and Analyst agents, and custom agents built in Agent Builder.
In April 2026, GA support for MCP arrived in Microsoft 365 Copilot declarative agents, making it easier for developers to integrate business workflows, SaaS systems and internal applications via the MCP protocol, without needing specific custom connectors.
To complete the picture, Microsoft developed its own MCP servers for M365 workloads, covering Outlook, Teams, SharePoint, OneDrive, Dataverse and Word, all respecting the same security, licensing and compliance boundaries that Microsoft 365 Copilot enforces internally.
Whoever Understands the Protocol, Controls the Agent
Let's leave the abstract aside. Some concrete consequences for those working in the M365 ecosystem:
Copilot Studio became much simpler to extend. Before, connecting an agent to an external system (an ERP, a CRM, an internal API) required a custom connector or Power Automate as an intermediary. With MCP, having a configured server is enough: actions and knowledge sources are automatically added to the agent as the MCP server evolves.
Declarative agents gain real access to external data. An HR agent can query the payroll system via MCP. A support agent can open tickets in Jira or ServiceNow. No connector code, no custom authentication to maintain.
Tool descriptions matter more than they seem. A real example: a maker discovered that their create_ticket tool had a description that matched close_ticket, causing the model to close tickets instead of opening them. The model uses the description to decide which tool to call. Knowing how to structure, name and describe endpoints correctly is what separates a functional implementation from one that makes production errors.
The Inflection Point Has Already Passed
In March 2026, Anthropic reported more than 97 million monthly SDK downloads and more than 10,000 active public MCP servers. MCP stopped being a trend and became a foundation.
For the M365 ecosystem, GA support in Copilot Studio and in declarative agents marks a clear turning point: integrating AI with external tools has stopped being niche work and has become a routine architecture decision.
If you're still evaluating whether it's worth understanding MCP, the market has already answered. The question now is when you start.
Subscribe to the Macareno.net newsletter and keep following what's changing in the Microsoft 365 ecosystem.
Sources
- Model Context Protocol is now GA in Microsoft Copilot Studio - Microsoft Copilot Blog
- Build declarative agents for Microsoft 365 Copilot with MCP - Microsoft 365 Developer Blog
- Agent 365 MCP Servers for Copilot Studio - The Custom Engine
- What is MCP: The 2026 Guide for SaaS PMs - Truto Blog
- One Year of Model Context Protocol - Ajeet Raina