Glossary
Definitions for the Model Context Protocol and MCP Risk vocabulary used across this site: MCP server, MCP client, tool, resource, prompt injection, trust grade, and capability risk. Each entry stands on its own.
- MCP server
- An MCP server is a program that exposes tools, resources, or prompts to an AI assistant over the Model Context Protocol, so the assistant can call functions, read data, or receive instructions beyond what it was trained on.
- MCP client
- An MCP client is the application, typically an AI assistant or an IDE plugin, that connects to one or more MCP servers and invokes the tools, resources, or prompts those servers expose.
- Tool
- In the Model Context Protocol, a tool is a named function an MCP server exposes that an AI assistant can call with structured arguments, such as running a shell command, querying a database, or sending an email.
- Resource
- In the Model Context Protocol, a resource is a piece of data, such as a file or a database record, that an MCP server makes available for an AI assistant to read. A resource is read; a tool is executed.
- Prompt injection
- Prompt injection is an attack where text an AI assistant reads, whether in a document, a web page, or an MCP tool’s own description, contains hidden instructions intended to override the assistant’s expected behavior.
- Trust grade
- The trust grade is an A-to-F rating, used by MCP Risk, of how an MCP server is built and operated: its provenance, transport security, known vulnerabilities, and instruction integrity. It does not rate what the server’s tools are capable of doing.
- Capability risk
- Capability risk is a Low, Medium, High, Critical, or Unknown rating, used by MCP Risk, of what an MCP server’s tools can do, such as running shell commands or reading credentials. It is reported rather than penalized: a server built to run shell commands is doing its job.