A new benchmark conducted by developer rabin-a exposes a critical inefficiency in the Model Context Protocol (MCP) proxy ecosystem. While many tools advertise significant token savings, real-world testing inside Claude Code revealed that four out of six popular proxies actually consumed more tokens than direct server connections. Two other proxies failed to complete tasks entirely, rendering them useless for production workflows. The findings prompted the developer to build mcp-router, a solution that achieved a 29% reduction in token usage across all six test tasks.

The Illusion of Savings in MCP Proxies

The core misconception driving the current MCP proxy market is the assumption that every tool schema is loaded upfront. In reality, Claude Code already optimizes for this by deferring tool loading unless explicitly required. Most proxies, however, introduce significant overhead by performing extra search or schema lookup turns that force the model to re-read the entire conversation history. This architectural flaw means that the savings advertised in marketing materials often translate into increased latency and higher token costs in practice. The benchmark specifically tested a fleet of 441 tools, a realistic stress test for modern agent architectures.

Benchmark Results: mcp-router Outperforms Competitors

The data presents a stark contrast between mcp-router and its competitors. While mcp-router reduced token usage by 29% and completed 6 of 6 tasks, lazy-mcp increased usage by 52%, and mcp-gateway (RaiAnsar) saw a 72% increase. Even worse, MCPGateway increased token usage by 145% and failed two tasks, while mcp-tool-router and mcp-context-proxy failed to complete any tasks due to search failures and result truncation errors. These results highlight that naive implementation of proxy layers can severely degrade performance, turning a cost-saving mechanism into a cost-center.

How mcp-router Achieves Efficiency

mcp-router operates by converging multiple tool connections into three standardized sockets: find_tools, call_tool, and run_script. This design allows Claude to search for specific tools on demand rather than carrying every schema in every request. The router ensures servers only start when a tool is first used, and it aggressively trims JSON results by removing link templates, nulls, and whitespace. This approach keeps large results inline and readable, avoiding the penalty of saving large outputs to files that require additional read-back operations. The tool is installed via a single curl command and supports both macOS and Linux on x86_64 and arm64 architectures.

Key Takeaways

  • Most advertised MCP token savings are theoretical and do not account for Claude Code's existing lazy-loading mechanisms.
  • Four of six tested proxies increased token usage, with some increasing costs by over 72%.
  • mcp-router reduced token usage by 29% by consolidating 441 tools into three efficient, on-demand interfaces.
  • Result size management is critical; keeping JSON outputs under 50,000 characters prevents costly file read-back operations.

The Bottom Line

Stop trusting marketing claims about MCP proxy savings. If your proxy doesn't account for Claude Code's native lazy-loading, it's likely just adding expensive overhead to your agent loops.