I've implemented a handful of tools for Anthropic Claude integrations at work: fetching public web page content, invoking Slack APIs, querying Amazon Bedrock Knowledge Bases. I've been hearing about Model Context Protocol (MCP) lately, and haven't had time to grok it's connection to LLM function calling and my experience.

With a good cup of ☕️ coffee and some time to read and think, I've arrived at the following understanding in brief.

---

• Many LLMs support function calling also known as tool use (e.g., OpenAI, Anthropic, Llama).
• There's no standard format for function call / tool use instructions across model vendors. Tool request and response formats vary.
• MCP "... provides a standardized framework for managing the execution of function call instructions, including tool discovery, invocation, and response handling"¹
• MCP servers host tools for discovery and execution.
• MCP clients bridge user interaction with LLMs, and LLM interactions with tools on MCP servers. For example:
  • User: "What's the weather?"
  • MCP client: List tools available for LLM use
  • LLM: Weather tool use request in LLM format
  • MCP client: JSON-RPC call to weather function
  • MCP server: Response to weather function call
  • MCP client: Weather tool use response in LLM format
  • LLM: "The weather is bright and sunny"
• MCP enables an ecosystem of adapters for existing services (e.g., https://mcp.so/).
• MCP uses JSON-RPC 2.0 on the wire with support for local (stdio) and remote (HTTP Server-Sent Events) transports.²
• It supports concepts beyond tools (e.g., resources, prompts).
• It fits into existing model broker architectures (e.g., Amazon Bedrock³, Ollama⁴).
• There are security, cost, development, hosting, pricing (and likely other) implications of the widespread adoption of MCP.⁵

References