RTK vs Headroom: Claude Code token usage compared

The speaker compares plain Claude Code with RTK, Headroom, and both together on the same subtitle feature. After twelve attempts across two evenings, one usable run per setup enters the comparison. Each starts from the same commit in a separate worktree, uses Opus 5 at medium effort, and stays within one agent.

RTK compresses command output; Headroom uses a local proxy to compress parts of model requests. Headroom also brings Serena through MCP for code navigation and editing. Setup problems included a lingering proxy environment variable and Serena configuration that initially listed only Python. The speaker added Swift and made Serena usage explicit. The final RTK comparison used version 0.49.0 after an earlier attempt with 0.46.0.

In these runs, RTK used 5.5% more tokens than the baseline. Headroom used 21.7% fewer, and combining the tools consumed more than Headroom alone. Totals include repeated cached input, planning, and the final coding answer. Headroom's estimated API-equivalent cost was only 9% lower. The speaker observed cache rewrites through the proxy, but the logs did not establish their cause.

This coding assistant comparison has limits: one run per setup cannot establish subscription savings, and the dollar estimates are not actual charges. All four implementations passed their own tests but failed an additional timeline-cut check. Repair work is excluded. The speaker keeps Headroom enabled for further use and turns RTK off for now.