chrlschn.dev
|
ksl
|
|
Charles Chen pushes back on the growing dismissal of Anthropic’s Model Context Protocol, arguing that critics are conflating two very different things. MCP over stdio for local agent work is arguably redundant – models already know tools like jq, curl, and git from training data. But MCP over HTTP solves a different problem entirely: centralized server deployment with OAuth, OpenTelemetry observability, and server-side secret management for organizations running agents at scale. The token savings debate, Chen argues, is mostly noise – the real value is security and control in multi-user environments. The piece is a direct response to influencers like Garry Tan and Andrew Ng who’ve been calling MCP unnecessary, and it lands at a moment when enterprise AI tooling is diverging sharply from solo developer workflows.
