When data is locked away, some build a closed loop inside their own walls (Alibaba), some try to climb over (Doubao Phone), and most just pretend the walls don’t exist. Who will push over the first brick?
The LLM is capable enough. The CLI is efficient enough. MCP is standardized enough. But try asking your agent to write your weekly status report — it can’t even read your Slack messages. The problem comes down to two distinct walls.
The hottest architecture debate in the AI Agent community in 2026. The CLI camp says MCP wastes tokens. The MCP camp says standardization is the future. The Skills camp says a single Markdown file is all you need. But after building an MCP Server myself, I realized the entire debate is dodging the real question.