MCP (Model Context Protocol)¶
Type: Protocol / standard
Definition¶
Model Context Protocol (MCP) is the emerging standard for exposing application APIs as tools that LLMs can call. It defines how an application describes its capabilities to an LLM, and how the LLM invokes those capabilities. MCP is similar to other web request frameworks (REST, GraphQL, gRPC) but is specifically designed for LLM tool use.
How It Works¶
- Application defines available tools (operations) via MCP
- LLM receives tool descriptions in its context
- LLM decides which tools to call based on user request
- Application executes the tool and returns results
- LLM incorporates results into its response
MCP enables both read operations (query application state) and write operations (modify application data). The key design question is which operations to expose — reads are generally safe, writes require safety mechanisms like version control.
Key Properties¶
- Standardized — common protocol across different LLM providers and applications
- Tool discovery — LLM learns what it can do from tool descriptions
- Bidirectional — supports both reads and writes
- Composable — tools can be combined for complex operations
- Similar to REST — familiar mental model for developers
Why It Matters¶
MCP is the glue that connects LLMs to applications. Without a standard protocol, every integration would be custom. MCP enables:
- LLMs to understand what actions are available
- Applications to expose capabilities in a structured way
- Ecosystem of reusable tool definitions
- The "chat on the side" pattern to work with any backend
Related Concepts¶
- cursor-for-everything — the pattern that uses MCP for tool exposure
- chat-on-the-side — the UI that presents MCP-powered capabilities
- branch-aware-llm-writes — safety for MCP write operations
- dolt-as-agentic-database — the database backend for MCP-powered apps
Source¶
- cursor-for-everything — full source article