Every MCP server you connect adds tool descriptions to the agent’s context — every turn, from the first prompt. Connect five servers and you can lose 30-50% of your context window before you’ve even asked a question.
This isn’t theoretical. LangChain research found that ~10 tools is the sweet spot for model accuracy. Beyond that, the agent gets worse at choosing the right tool — the menu is too long.
Toolmakers are noticing. Claude Code recently shipped “Tool Search” — a lazy-loading system that only loads tool definitions on-demand. Anthropic reports 85-95% reduction in initial context usage. It activates automatically. Smart, but currently Claude Code-specific.
In Copilot CLI, the pattern is: be selective. Type /context to see your token budget. The GitHub MCP server and Chrome DevTools MCP (Tip #13) earn their context cost. But “install every MCP server that looks cool” is a trap.
And remember: if a CLI tool can do the same job (Tip #14), it costs zero context until the agent calls --help.
💡 Try this: Run /context in Copilot CLI. How much of your budget is tool descriptions vs. actual conversation?
🔗 Claude Code Tool Search (search “ToolSearch”)
🔗 GitHub docs on MCP toolsets
How many MCP servers do you have connected?
🟢 Just the GitHub one
🟡 2-3 carefully chosen
🔴 Lost count — maybe too many
⚪ None yet, still figuring it out
Tips for keeping your context lean?