After 100+ team interviews, Swytchcode argues MCP is for building agents, not running them
curiousblack99 · reddit · 2026-11-05
After interviewing 100+ teams shipping agents against third-party APIs (Oracle, Salesforce, Stripe, Deepgram), the author found a consistent pattern: MCP speeds up building but hurts in production — tool schemas eat context every turn, the model handles auth refreshes/retries/schema drift alone, and there's no clean policy boundary (e.g. "read invoices but never issue refunds").
Swytchcode splits the two phases:
- Build with MCP: ships an MCP server covering 350+ APIs / 100k+ methods, generating production-grade integration code so agents get the right calls in one shot.
- Run with the runtime: production calls go to a self-hosted runtime doing schema validation, auth, allow/deny/approval policies, retries + idempotency, and logging — no tool schemas in the model's context.
Headless support includes Hermes, OpenClaw, OpenAI Agents SDK, Anthropic SDK, Vercel AI SDK, LangGraph, and CrewAI. Honest caveats: the runtime runs in your infra (CLI + locally encrypted credential store), but the control plane (OAuth brokering, catalog, telemetry) is managed cloud that never sees API payloads; the CLI isn't open source; there's a free tier.
More from coding & agent
- Engineer says manual tweaks after Claude optimization made EC2 CPU 50x faster than M4 Max — _Stocko_ · 2026-11-05
- A Rust todo app taking 14GB? Compile caches, not the app, eat ~13GB — lxfater · 2026-11-05
- Remote MCP: managing your devices remotely via the MCP protocol — Latter-Storage9610 · 2026-11-05
- MCP over HTTP: validating OAuth tokens isn't enough—check the audience claim — alvelda · 2026-11-05
- Short RL phase on GLM 5.3 Flash costs $372 vs $1,894 for notes, breaks even in 8.7 days — teortaxesTex · 2026-11-05
- Vercel Security Review finds 6,817 vulnerabilities in 24 hours, agent patches 1,411 — cramforce · 2026-11-05