Cite as: Real Problem AI problem “Why does connecting one large MCP server cost context on every single turn, even when you never touch most of its tools?”. Opportunity score 6.5 out of 10 (severity 6, AI feasibility 8, market signal 5, competition gap 7). Category AI / Agents. Trend Agents. Source signal: anthropics/claude-code GitHub issue #27208, DanielPBak, 20 Feb 2026, closed as duplicate.. Canonical URL: https://www.realproblem.ai/archive/why-does-a-200-tool-mcp-server-cost-context-on-every-single-turn.
Why does connecting one large MCP server cost context on every single turn, even when you never touch most of its tools?
A gRPC-proxy MCP server exposing ~200 tools across 8 services has every tool name listed in the system prompt on every turn for ToolSearch, even though full schemas are deferred and most tools are never used in a given session.
Who has it: Developers connecting large enterprise MCP servers to Claude Code.
Evidence
The issue describes how every deferred tool name from a large MCP server is still listed in the system prompt, so the name list alone uses context on every turn.
Our summary of the public post linked below, not a quote. Nobody submitted it to Real Problem AI.
anthropics/claude-code GitHub issue #27208, DanielPBak, 20 Feb 2026, closed as duplicate.Why it is archived
Trimmed to 100-cap (lowest opportunity_score)
Scoring breakdown
Existing players
- Deferred tool loading · Reduces schema cost but still lists every flat tool name up front.
- Manual server disconnection · Avoids the cost entirely but loses the server's functionality.
What they are missing
Hierarchical/nested tool discovery where the system prompt shows only service-group names, and ToolSearch drills into individual tools and schemas on demand, cutting a 200-entry list down to ~8.
Stack hint
#ATB29 · Canonical URL: https://www.realproblem.ai/archive/why-does-a-200-tool-mcp-server-cost-context-on-every-single-turn