What is MCP (Model Context Protocol)?
The Model Context Protocol — an open standard for connecting an agent to external tools and data, so a system exposes its capabilities once instead of once per AI product.
MCP is a plug standard. Before it, connecting an agent to your CRM meant writing an integration for that agent and that CRM; connecting a second agent meant writing it again.
With MCP, a system exposes an MCP server describing what it can do, and any MCP-speaking agent can use it. The N-times-M problem becomes N-plus-M.
What it actually looks like
An MCP server publishes a list of tools — each with a name, a description of when to use it, and the shape of its inputs. The agent reads that list, decides a tool is relevant, calls it with arguments, and gets a result back to reason about.
Nothing about it is AI-specific. It is closer to a typed API description written for a reader who decides at runtime which endpoint it needs, rather than a developer deciding at build time.
The kind of thing that belongs behind a tool
Affordability. Points-based eligibility. Premium loading. Anything where the rules are long, the answer must be exactly right, and someone may later have to show how it was arrived at.
A mortgage quote is the clean example. The rules run to pages, they change when regulation changes, and a number that is wrong by a hundred is not a wording problem — it is a complaint. So the calculation stays in a function your own systems already use, exposed as a tool, and the agent's job is to gather the inputs, call it, and report what came back.
What that leaves behind is auditable: these inputs went in, this number came out. That is a very different artefact from a paragraph of plausible reasoning that happened to end in a figure.
Why it matters commercially
Two things follow from an open standard, and both are about not being trapped:
Your integrations outlive your model choice. A tool exposed over MCP keeps working when the underlying model changes, because the contract is with the protocol rather than the vendor.
Your systems stay yours. An MCP server runs where you run it, exposing exactly the tools you choose to expose. Connecting an agent to your booking system does not mean handing your booking system to anyone.
How tools are reached here
An agent does not simply get handed every tool the tenant has connected. Tools are enabled by the Skills an agent carries, and a Skill loads only when the conversation calls for it — so a tool no loaded Skill enables is a tool the agent will not reach for.
There is also a guard for the failure mode that matters most commercially. An agent can be set to require grounding: before it states a specific figure it must have actually called a tool or search to obtain it. If it produces a number without having done so, the turn is re-run. That converts "the model usually looks things up" into something closer to a rule.
MCP runs in both directions here. Your agent can call your systems through it — and these docs are themselves served over MCP, so a coding agent can read them and wire up an integration without you transcribing anything. See Set it up with your coding agent.