Insight · July 24, 2026
One protocol
for every tool.
Model Context Protocol is the standard that lets any AI agent reach any tool through one connection, instead of a hundred custom ones.
01 · The problem
Every tool spoke a different language.
Before MCP, connecting a model to your systems meant building a custom bridge for each one. Your CRM had its own integration. Your database had another. Your ticketing system, your calendar, your files, each needed its own glue code, written and maintained by hand.
The math got ugly fast. Ten tools and three models is thirty integrations, and every one of them can break on its own. Try a new model and you rebuild all of it. Teams spent more time wiring things together than using what they built.
02 · What it is
One open standard, spoken by both sides.
Anthropic released the Model Context Protocol in November 2024 as an open standard. The idea is plain. Instead of a custom bridge for every pairing, there is one shared language. An agent that speaks MCP can talk to any tool that speaks MCP, with no new code on either side.
Underneath, it is a client and a server exchanging structured messages in the JSON RPC format. The agent runs a client. Each tool or data source runs a small server. The server describes what it can do, the client asks, and the work happens. That is the whole shape of it.
People reach for the same comparison every time. MCP is the USB C of AI tools. One port, and everything plugs into it. You built the port once, and every agent that comes later can use it.
The one line to keep
“MCP is not a feature you bolt onto a model. It is the port every tool now plugs into.”
03 · What a server offers
Three primitives, and everything is built from them.
01
Tools
Actions the agent can take. Send the email, run the query, create the ticket. The server names each one and says what it needs to run it.
02
Resources
Data the agent can read. A file, a record, a page of results. Context pulled in on demand, not stuffed into the prompt ahead of time.
03
Prompts
Reusable templates the server offers. A tested way to ask for a common job, so the agent does not have to invent the phrasing every time.
04 · Not just an API
An API waits to be called. MCP describes itself.
A normal API sits there until a programmer reads the docs, writes the calls, and ships them. Nothing happens without a person in the middle translating intent into requests.
MCP flips that. The server publishes what it offers in a form the agent can read at run time. The agent discovers the tools, understands what each one needs, and decides which to use. You are not writing the integration for every task. You are exposing a capability once and letting the agent choose when to use it.
That is why it caught on the moment agents did. An agent is only as useful as the things it can reach. MCP is how it reaches them.
05 · What it looks like in practice
Wire one tool once. Every agent gets it.
Say you want your support agent to look up orders. Without MCP, you write a bespoke integration inside that one agent, then do it again for the next agent, and again when you switch models.
With MCP, you stand up one small server that exposes a single tool. Any agent that speaks the protocol can now use it, today and next year, whichever model sits behind it. The description below is all the agent needs to find the tool and call it correctly.
{
"name": "lookup_order",
"description": "Find an order by its number and
return its status and items",
"inputSchema": {
"type": "object",
"properties": {
"order_number": { "type": "string" }
},
"required": ["order_number"]
}
}One tool definition · read and called by any MCP client
06 · Why it became the standard
The whole industry stopped arguing and agreed.
Standards usually take years and a fight. This one did not. OpenAI adopted MCP across ChatGPT and its developer platform in March 2025. Google confirmed support for Gemini weeks later. Microsoft built it into its agent and Microsoft 365 tooling. GitHub and Amazon shipped their own servers.
In September 2025 an official registry launched so agents could find servers in one place. It passed close to two thousand entries within months. In December 2025 Anthropic handed the protocol to the Agentic AI Foundation under the Linux Foundation, so no single company owns it.
The 2026 specification, finalized on July 28, made the protocol stateless so servers scale on ordinary web infrastructure, and added interactive interfaces and long running tasks. For a business, the signal is simple. This is not a bet on one vendor. It is plumbing the whole field agreed to share.
07 · The part to get right
A tool the agent can reach is a tool an attacker can aim.
The same openness that makes MCP useful makes it a surface to defend. A server describes its own tools, and the agent trusts that description. If a malicious server hides instructions inside a tool description, the agent can read and act on them while the user sees nothing. Researchers call it tool poisoning, and it works on every call until someone notices.
There is a second risk in the data itself. When an agent reads a resource, that content can carry instructions too. A tampered document or web page can try to steer the agent, a class of attack known as indirect prompt injection.
None of this is a reason to avoid MCP. It is a reason to run it like production. Give each server the least access it needs, keep a short list of servers you actually trust, and put a person in front of anything destructive, anything that moves money, or anything that leaves your systems. The protocol connects the tools. The discipline is still yours.
Closing
Your software was never meant to be an island.
Pick one system your team reaches for every day. Put an MCP server in front of it. The first time an agent uses it without anyone writing a new integration, you will stop building bridges one at a time.
Model Context Protocol announcements and specification · modelcontextprotocol.io, November 2024 through July 2026 · adoption timeline from public releases by Anthropic, OpenAI, Google, and Microsoft
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.