MCP is dead

MCP (Model Context Protocol) has drawn attention as a standard for connecting LLMs with external tools, but technical limitations arising in actual operation have been pointed out.

1. Excessive occupation of the context window The tool definitions used when connecting MCP servers consume a large amount of the LLM's context window. Experimental results showed that when 4 MCP servers were connected, 10.5% of context was used for tool definitions on Claude (200K), and 16.5% on GPT-4o (128K). In particular, a single Linear server alone occupied about 12,800 tokens, showing that the share of certain tools was very high.

2. Low operational reliability and performance degradation Because MCP places an additional process layer between the LLM and the API, performance degradation occurs. When comparing the Jira MCP to the REST API, it showed 9.4x slower performance on the first call, and 3x slower on regular calls. In addition, there are operational issues such as process crashes, the need for re-authentication, and opaque permission management.

3. Overlap with and closedness compared to existing CLI/API MCP makes it difficult to freely combine tools using pipes (|), jq, grep, etc., as with existing CLI tools, and it is dependent on the tool's return format. It also has a closed structure that only works within LLM conversations.

Recently, Claude Code's 'Tool Search with Deferred Loading' feature has partially mitigated the problem by reducing context usage by more than 85%, but architectural and performance issues still remain challenges to be solved.