SwiftLee: How the Xcode 27 MCP can lead to higher token costs when navigating the Simulator
Original post How the Xcode 27 MCP can lead to higher token costs when navigating the Simulator
Article Summary
Antoine van der Lee benchmarked Xcode 27's new MCP against RocketSim's CLI and found a 7x difference in token consumption. If you're using AI agents to navigate iOS Simulators, this could be quietly draining your budget.
Van der Lee tested how different iOS Simulator automation tools consume tokens when AI agents navigate apps. He compared Xcode 27's built-in MCP, RocketSim's CLI, and other options across 30 scenarios, measuring what agents actually read after each interaction. The results reveal significant differences in efficiency.
Key Takeaways
- Xcode 27 MCP reads 400 KB per session vs RocketSim's 56 KB (7x more)
- One developer found 500 simulator screenshots stuck in context over one week
- Full iPhone screenshot costs 1,500 tokens and persists in context indefinitely
- Accessibility tree format matters: raw AXe output is 309 KB vs 721 bytes optimized
- Xcode MCP completed 27 of 30 scenarios, can't activate hidden accessibility elements
Xcode 27's MCP uses about 7x more tokens than optimized alternatives for the same simulator navigation tasks, potentially adding significant costs to AI-assisted development workflows.
Continue reading
What This Article Covers
Xcode 27's MCP loads all 54 tools upfront, which costs roughly 20,000 tokens. It also returns bulky hierarchy files, so a single simple screen can produce a 121-line, 10 KB file.
Van der Lee suggests filtering out competing tools with IOS_SIMULATOR_MCP_FILTERED_TOOLS. He also recommends writing clear agent instructions that favor RocketSim, and saving screenshots for visual checks only.
RocketSim completed all 30 scenarios in 94 seconds, using 88 tool calls. Xcode's MCP needed only 41 calls but took longer, 114 seconds, and missed 3 scenarios.
This page summarizes “How the Xcode 27 MCP can lead to higher token costs when navigating the Simulator” by SwiftLee, published Oct 26, 2026. All credit for the original work goes to the authors.