MCP and Tool Governance

Model Context Protocol (MCP) is an open standard for connecting AI agents to external tools, data sources, and services through a structured client-server protocol with defined methods for tool discovery and invocation.

An MCP server is a service that exposes tools, resources, and capabilities to AI agents via the Model Context Protocol, responding to tools/list and tools/call requests.

An MCP client is the component within an AI agent or agent framework that initiates connections to MCP servers, sends tools/list requests to discover available tools, and sends tools/call requests to invoke them.

An MCP proxy is an intermediary that sits between an MCP client and an MCP server, able to inspect, filter, modify, allow, or deny tool calls before they reach the underlying server.

An embedded MCP proxy is an MCP proxy deployed within the agent's own execution environment rather than as a separate network service, providing enforcement close to the tool call with minimal latency.

MCP tool discovery is the process by which an AI agent queries an MCP server using the tools/list method to enumerate the tools the server exposes, their schemas, and descriptions.

tools/list is the MCP protocol method that an agent client sends to an MCP server to retrieve the list of available tools, their input schemas, and descriptions.

tools/call is the MCP protocol method that an agent client sends to an MCP server to invoke a specific tool by name, passing input parameters defined in the tool's schema.

A tool schema is the structured definition of a tool's name, description, input parameters, and types exposed via MCP, used by agents to understand how to invoke the tool correctly.

A tool catalog is the collection of tools available to an AI agent across one or more MCP servers, representing the full action surface the agent can potentially invoke.

Tool catalog filtering is the practice of restricting the tools an agent can discover and invoke by returning a filtered tools/list response based on the agent's identity, task context, and access policy.

Tool invocation is the act of an AI agent sending a tools/call request to execute a specific tool, triggering a real-world action such as reading data, calling an API, writing a file, or executing code.

The tool-call boundary is the enforcement point between an AI agent's decision to invoke a tool and the tool's execution, where access policy can be evaluated, credentials can be issued, and calls can be allowed or denied.

Tool-call authorization is the process of deciding, at the moment an agent attempts to invoke a tool, whether that invocation is permitted based on the agent's identity, task context, policy, and intent.