Glossary term
MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard that defines how an AI application connects to external tools, data sources, and prompts over a uniform client–server interface, so that any compliant client can use any compliant server without integration code written for that specific pair.
Also known as Model Context ProtocolMCP serverMCP client
In practice
The protocol specifies a JSON-RPC exchange between a client and a server. A server advertises tools (callable operations), resources (readable data), and prompts (reusable instruction templates); a client — a chat application, an IDE, or an agent runtime — discovers them when it connects and calls them over stdio or HTTP. It was introduced by Anthropic in late 2024, published as an open specification, and is now implemented by clients and servers from many vendors and open-source projects. The value it captures is combinatorial: N clients and M tools stop requiring N×M bespoke integrations.
A connection standard says nothing about authority, and MCP deliberately does not try to. It describes how a call is framed and transported; it does not decide whose permissions apply to that call, whether the operation needs a signature before it takes effect, or what evidence remains afterwards. Those decisions belong to the system holding the data, which is why an MCP server is never safer than the interface it wraps: a server placed in front of a raw database connection hands the agent the service account reach, not the reach of the person asking.
The clean way to hold the two apart is that MCP standardizes how an agent reaches a tool, while a governed tool layer decides whether the call is allowed and whether it is recorded. Both are needed. Connectivity arriving years before governance is exactly how an agent ends up cheerfully exporting records that the person who asked was never entitled to see — not through malice, but because nothing in the path ever asked who was calling.
Where this term is used
Pages and articles that put this term to work.
Product pages
- Tools & MCP The @objectstack/mcp server turns objects, queries, and actions into policy-checked tools for Claude, Cursor, or any MCP client — with permissions enforced on every call.
- AI Build & Ask ObjectOS AI Build turns a described change into reviewable ObjectStack metadata. AI Ask answers over live business data inside user permissions; open-source ObjectStack supports your own agent via MCP.
Articles
- MCP Security for Enterprise Agents: Why Protocols Need Governed Tools MCP can connect agents to tools quickly, but enterprise systems need identity, permissions, approvals, and audit behind every tool call.
- What Is Ontology MCP? Turning an Enterprise Ontology Into Agent Tools Ontology MCP exposes an ontology's object types, action types and functions as MCP tools any agent can call. What Palantir shipped, what Microsoft is previewing, and the tool-surface asymmetry to plan for.
- Ontology Actions: How to Govern AI Agent Write Operations Enterprise ontology work stops at retrieval. When an agent has to change state, write operations need one governed definition — opt-in exposure, permission parity, confirmation and audit — not a second tool path beside your UI.
- How AI Agents Stay Inside Enterprise Permission Boundaries Enterprise teams do not need AI agents to become unrestricted administrators. They need agents that act as controlled users, inherit permissions, route risky actions for approval, and leave an audit trail.