MCP tools consume 50% of context tokens - need consolidation strategies

Problem

MCP tools currently consume 98.7k tokens (49.3%) of my 200k context window, leaving limited space for actual work.

Current breakdown:

• System prompt: 3.2k (1.6%)
• System tools: 16.1k (8.0%)
• MCP tools: 98.7k (49.3%)
• Memory files: 3.0k (1.5%)
• Free space: 34k (17.0%)
• Autocompact buffer: 45k (22.5%)

MCP Servers Connected

| Server | Estimated Tokens | Tools |
|--------|-----------------|-------|
| google-docs | ~25k | 34 tools |
| asana | ~30k | 40+ tools |
| google-calendar | ~15k | 12 tools |
| slack | ~3k | 4 tools |
| atlassian | ~12k | 15 tools |
| gmail | ~8k | 9 tools |

Questions

1. Selective tool loading: Can MCP servers expose only a subset of tools based on config? (e.g., load only search_tasks, get_task, create_task from Asana instead of all 40+)

1. On-demand tool discovery: Could tools be loaded dynamically when needed rather than all upfront?

1. Tool description compression: Are there best practices for MCP server authors to minimize tool description token usage while preserving functionality?

1. Per-session MCP selection: Can I configure which MCP servers load for specific project types? (e.g., skip google-docs when doing pure code work)

1. Consolidation patterns: Has anyone built "meta" MCP servers that wrap multiple services with a unified, smaller interface?

Desired Outcome

Reduce MCP token usage to ~25-30% of context while maintaining access to core functionality across all connected services.

Environment

• Claude Code CLI (latest)
• macOS
• 6 MCP servers connected