Brief, mostly signal

MCP servers eat 56% of context before a session starts

A Reddit question asks for lazy loading of MCP servers. No tool or answer is named, so only the problem is confirmed.

What the research found

A user asks whether a lazy-loading layer exists so many always-relevant MCP servers do not consume 56% of context at session start; the post is a question with no answer or tool named, so only the problem is confirmed.

Claims and how well the evidence supports them

A new Claude Code session starts with 56% of the context window already used by MCP servers 95%
The user has many MCP servers that are all relevant to daily work 99%
A per-MCP lazy loading layer exists for Claude Code (jev unsupported, post names no tool) 11%

Evidence

reddit.com https://reddit.com/r/ClaudeAI/comments/1xayhwt/lots_of_mcps_new_sessions_start_with_56_of

Actionability rated low.