passdrill
AI & LLM Engineering · AI Agents & Tool Use · Card 014/022 easy

Before adopting the Model Context Protocol (MCP), a company had built a separate bespoke integration connecting its coding-assistant agent to each of a Git server, a database, and an internal ticketing API -- three different one-off connectors, each written specifically for that one agent and that one system. MCP's host-client-server architecture instead has the host application create one client per server, with each client keeping a connection to exactly one server. What problem is this specifically designed to solve?

  1. It replaces separate one-off integrations, one per agent-to-tool pairing, with a shared protocol that any MCP-compatible host can speak to any MCP-compatible server, so a new tool only needs to be exposed once as a server rather than re-integrated for every agent that wants to use it
  2. It removes the need for any security boundary between the agent and the systems it connects to, since MCP's stateless request format already guarantees every request is safe to run without further checks
  3. It requires that a single MCP server be shared across every client in the host, so that all the tools from the Git server, the database, and the ticketing API are merged into one server implementation
  4. It guarantees the coding-assistant agent will never need to be given more than one tool at a time, since MCP only allows a single tool call to be issued per client
Next card → Shuffle