Stacking MCP servers silently degrades AI agent performance — no way to scope tools per task

Summary

Users building on agents stack up MCP servers (GitHub, database, browser, Figma, docs) and then report the output getting 'a lot worse, not better,' context limits hit almost immediately, and even broken projects — because every server injects more tool descriptions the agent must sort through. They are forced into a false choice: leave everything connected 'just in case' and degrade, or manually toggle servers off per project and lose the promised leverage. This is an unsolved, increasingly painful problem for exactly the power users and agent builders this topic targets: there is no tooling that profiles which MCP servers actually help a given task, no automatic context budgeting, no per-task tool scoping, and no diagnostic telling the user which server is poisoning the run. The failure is silent and misattributed to the model, so people burn hours debugging prompts while the real culprit is tool sprawl. As MCP adoption grows this becomes a mainstream blocker worth paying to solve — a 'context/tool governor' for agents mimics the abstraction everyone is hand-rolling today.