The case against MCP: autonomous agents collapse back to APIs and scoped tokens

uriwa · reddit · 2026-11-19

A contrarian long-form argument that MCP is a transitional layer built for human-driven IDE/desktop copilots. For autonomous agents running on their own runtime (webhooks, cron jobs), it collapses into two older primitives: documented APIs and scoped, revocable tokens. Four points: (1) modern models read OpenAPI specs/docs/llms.txt and write working integrations, so per-service MCP servers are a drifting middle layer; (2) AI writes both the docs and the MCP manifests, so MCP schemas aren't inherently cleaner; (3) hiding raw API keys in a local MCP sidecar isn't a real security boundary when you already trust the agent — put authorization at the token/runtime layer; (4) sandboxed scripts that call APIs directly beat stuffing megabytes of raw JSON through tool calls. Author asks where MCP remains irreplaceable for background agents.

Related event: Autonomous agents may ditch MCP for APIs and scoped tokens(2 posts)→

Original post →

More from coding & agent

coding & agent channel →