Model Context Protocol
The Model Context Protocol (MCP) is an open protocol for connecting AI applications to external data and capabilities through shared client–server contracts. It standardizes an integration boundary; it does not itself decide which context is relevant or which action is authorized. Anthropic MCP announcement · Tools and interoperability source record
Motivation
Anthropic’s 2024 launch announcement frames MCP as an answer to repeated custom connectors between assistants and data systems. A provider exposes data or capabilities through an MCP server, while an AI application connects as a client. Anthropic MCP announcement
The later Google whitepaper places MCP at the model-or-host-to-tool boundary and contrasts it with A2A, which coordinates autonomous peers. Tools and interoperability source record (pp. 6–17)
Architectural role
MCP can support:
- capability discovery through machine-readable definitions;
- retrieval of context from external systems;
- invocation of tools that read or change state;
- reuse of a server implementation across compatible clients.
Analysis: MCP is best understood as part of an agent harness. The harness still chooses servers, assembles context, scopes credentials, validates outputs, records traces, and decides whether a model may invoke a capability.
Security boundary
A common interface can increase reuse and also increase the impact of a compromised or misleading server. Important controls include:
- explicit server trust and provenance;
- least-privilege credentials;
- user consent for sensitive operations;
- schema and output validation;
- separation of read and write capabilities;
- network and sandbox boundaries;
- logging, revocation, and incident response.
The launch article calls MCP secure but does not establish that every implementation satisfies these controls. Anthropic MCP announcement · Agent safety and control
Evolution and versioning
The Anthropic article is a launch snapshot, not a current specification. The Google whitepaper also describes one stage of a changing ecosystem. Implementations should pin a protocol version and verify transport, authorization, lifecycle, and compatibility behavior against that version.
See MCP vs A2A for boundary and interaction differences.
Evidence status
Confidence is high that MCP originated as a shared client–server integration protocol and medium for broad claims that it reduces lifecycle cost or guarantees portability. Those benefits depend on conformance, semantics, trust, and independent implementations.