The Model Context Protocol (MCP) lets developers expose tools and data sources to LLM-based agents through a standardized interface. Despite rapid ecosystem growth, no methodology exists for evaluating whether a given MCP server improves agent task completion. We present mcpbr, an open-source benchmark runner that isolates the effect of MCP tool augmentation through paired comparison experiments. We evaluate a code graph analysis MCP server on all 500 tasks from SWE-bench Verified using Claude Sonnet as the base agent. MCP augmentation reduced resolution rate by 14.9% (from 49.8% to 42.4%) while improving efficiency: 42.3% fewer tool calls, 14.0% fewer tokens, and 15.2% lower cost. Per-repository analysis shows the effect varies across codebases, with the server helping on 1 of 12 repositories and hurting on 10. We analyze this efficiency-resolution tradeoff and show that MCP tools alter the agent's exploration strategy, trading general-purpose search for opinionated shortcuts that can narrow the solution space.
Grey Newell (Thu,) studied this question.