MCP (Model Context Protocol)
The Model Context Protocol (MCP) is an open standard that connects AI assistants and agents to external tools, data sources and services through a common interface. It lets an assistant like Claude call live tools (such as a security data server) instead of relying only on its training data.
MCP standardises how AI clients discover and call tools, so any MCP-compatible assistant (Claude, Cursor, Zed and others) can use any MCP server without bespoke integration.
For security, this means an agent can fetch a live CVE, its EPSS score and KEV status mid-conversation, and reason with current data rather than a stale snapshot.
MCP was introduced to solve a fragmentation problem. Before a shared standard, every integration between an AI application and an external tool or data source was bespoke, which did not scale as the number of models and tools grew. MCP provides a common protocol so that any compliant client can talk to any compliant server, in the way a universal port replaced a drawer full of proprietary connectors.
The protocol follows a client-server model. An MCP host, such as an AI assistant or development tool, runs one or more clients that connect to MCP servers. Each server exposes capabilities in standard categories: tools that the model can call to take an action or fetch data, resources that provide context, and prompts that offer reusable templates. The model discovers and uses these through the protocol without custom integration code.
MCP matters because it turns isolated language models into agents that can act on current, external information. A model on its own is limited to its training data; connected through MCP to live tools and data sources, it can query systems, retrieve up-to-date information and perform tasks. This is the foundation of the agentic AI applications that have driven the protocol’s rapid adoption.
Because MCP servers can expose real capabilities, security is a first-class concern. Authentication, authorisation, careful control over what a server is allowed to do, and awareness of risks such as prompt injection through returned data all matter when connecting a model to tools. A well-designed MCP server makes its access model explicit, which is why read-only, clearly scoped servers are a safer starting point for security use cases.
MCP was introduced by Anthropic in late 2024 and released as an open standard, and it saw rapid adoption across AI tools and platforms. That momentum matters because the value of a protocol grows with the number of clients and servers that speak it: a widely adopted standard means a single MCP server can serve many different AI applications, and a single application can reach many tools, without bespoke work for each pairing.
MCP servers communicate with clients over defined transports, commonly a streamable HTTP transport for remote servers or a local process transport, using structured JSON-RPC messages. This standardisation is what lets an ecosystem form: developers build a server once, against the protocol, and any compliant assistant, whether a desktop application, a coding tool or a custom agent, can discover and use its tools, resources and prompts.
The Precursor MCP server exposes 41 security tools and unifies CVE, EPSS, CISA KEV, threat-actor and Shadowserver data behind the exploitation risk score.