Skip to main content
MCP (Model Context Protocol) is an open protocol created by Anthropic that standardizes communication between AI agents and external tool servers. Instead of each platform inventing its own integration format, MCP defines a common contract: any agent that implements the protocol as a client knows how to talk to any server that implements it as a server — regardless of what that server offers or what language it was developed in.

​
Why MCP matters

Before MCP, integrating tools into an AI agent was manual and fragmented work: each platform had its own tool calling format, and the same tool server had to be rewritten to work with different systems. MCP eliminates this fragmentation with a single protocol.
  • An MCP server developed once works with any compatible client — the Timely.ai agent, Claude Desktop, code editors with AI support, and others.
  • The agent automatically discovers the tools available on the server via the protocol — no manual documentation, no static mapping.
  • Public MCP servers developed by the community can be connected directly to your agent without modification.
  • Proprietary tools (internal APIs, ERPs, legacy systems) become accessible to the agent without exposing a public REST API.

​
MCPs in action

In practice, MCP works as a universal interface between the agent and any external system you want to connect:
  • An MCP server with specific queries to your internal database — the agent queries records without needing a dedicated REST integration.
  • An MCP server that aggregates multiple internal APIs and exposes a simplified interface to the agent — the agent calls one tool and the server orchestrates the necessary calls.
  • A public MCP server for web search, file management, or browser control — connected to the agent with no adaptation required.
The Timely.ai agent acts as an MCP client: it maintains a session with each configured server, discovers the available tools, and registers them in the model context. When the LLM decides to trigger a server tool, the client executes the call via SSE or HTTP, receives the result, and injects it into the context before the next iteration.

​
Key characteristics

MCP was designed specifically for use by AI agents — and this is reflected in its fundamental characteristics:
  • Automatic tool discovery — the agent reads the tool schema directly from the server, without manual intervention.
  • Communication via SSE or HTTP — compatible with any language and server infrastructure.
  • Stateful session support — the protocol allows the server to maintain context between calls in the same session.
  • Model-readable descriptions — each tool exposes a name, description, and parameter schema that the LLM uses to decide when and how to trigger it.