Should stdio MCP servers be launched only when a tool is called?
gelembjuk · reddit · 2026-07-24
This is essentially the same proposal as the earlier post: run stdio MCP servers only when a tool call needs them, rather than keeping them alive all the time.
The motivation is the same too: most MCP servers are stateless API wrappers, so always-on processes waste RAM. The post asks whether harness libraries should change to support this on-demand lifecycle.
Related event: On-Demand MCP Server Startup Saves Hundreds of MB of RAM(2 posts)→
More from coding & agent
- LLM security tooling looks dramatically better in the Sonnet 4.7+ era — dsp_ · 2026-07-24
- NO8D adds 397 prompt cards and Krea 2 styles to its ComfyUI library pack — Suspicious_Aide2697 · 2026-07-24
- Why LLM agents need four hard boundaries before they can touch production — Innowise_ · 2026-07-24
- Google and Gemini Notebook workflows are becoming skills, but mostly for reliability and latency — tokumin · 2026-07-24
- Building Profit-Driving Agent APIs is the Current Goldmine — eptwts · 2026-07-24
- Grok Build adds workflows that fan tasks out across hundreds of agents — elonmusk · 2026-07-24