Six token-reduction tools hit 200,000 stars with no shared benchmark

More than 200,000 GitHub stars now back six tools built to cut AI agents' token costs, per GitHub. Each claims a different reduction number, from 20 percent to 99 percent. None of the six publish a shared benchmark, so a developer choosing between them is picking on star count, not evidence.

rtk-ai/rtk leads with 74,640 stars, a Rust binary claiming 60 to 90 percent reduction on dev commands. headroomlabs-ai/headroom sits at 64,556 stars and splits its number: 20 percent for coding agents, 60 to 95 percent for JSON. DeusData/codebase-memory-mcp claims 99 percent reduction at 37,389 stars. tirth8205/code-review-graph has 28,384 stars, mksglu/context-mode claims 98 percent tool-output reduction at 19,606 stars, and yvgude/lean-ctx trails at 3,486.

All six repos were created in 2026 and pushed within 48 hours of each other. Each works the same way, a local binary between the agent and its tool output, filtering what reaches the model. But a 90 percent claim on dev commands and a 99 percent claim on an unnamed workload aren't the same measurement. They're different tests wearing the same unit. Nobody in the category has published one the other five could run against.