MCP was always a bad idea?

The Model Context Protocol (MCP) has come under scrutiny as users find it outdated and inefficient in light of advancements in large language models (LLMs). Launched in November 2024 by Anthropic, MCP was designed to help agents connect to external services and data sources, but it was released before the current level of LLM sophistication.

As LLMs have grown in capability, users have seen the benefits of granting their AI models direct access to external services, leading to an explosion in MCP adoption. However, the protocol's growing popularity has led to context bloat, as each server comes with multiple tools and schemas, overwhelming model context. To address this issue, developers have found workarounds using generic search/execute patterns offered by platforms like Composio, MintMCP, and Pipedream, which reduce context bloat by providing agents with a minimal set of tools to access external services.

Despite these solutions, the evolution of LLMs has rendered many MCP servers unnecessary. With LLMs now able to execute code, reason about large codebases, and call APIs directly, the need for MCP servers has diminished. Some remote-service MCP servers wrap APIs that already exist, and the majority of MCP servers can be deleted, as agents with terminal access can replace most MCP servers and are often more capable.

Standardizing HTTP API usage for agents, such as attaching headers to identify themselves and requesting content in Markdown format, could further streamline agent interactions with external services. With the technology landscape continually evolving, it is time to leave the MCP protocol in the past and rely on HTTP APIs and CLIs for direct agent-agent communication.