RTK (Rust Token Killer)

What It Is (confidence: medium)

RTK (“Rust Token Killer”) is an open-source (Apache 2.0, Rust) CLI proxy that intercepts shell commands run by an AI coding agent and compresses their output before it reaches the model’s context window. Single Rust binary, ~74.1k stars at capture, <10ms overhead, 100+ supported commands. Works with Claude Code, Cursor, Aider, Gemini CLI, OpenAI Codex, Cline, Windsurf, and GitHub Copilot.

Claims (confidence: medium)

• Compresses command output before it enters context: cargo test 91.8% noise removed, git status 80.8%, find 78.3%, grep 49.5%, ~89% average across 2,900+ measured commands.
• Token numbers are estimated as bytes / 4 — RTK ships no tokenizer, so percentages are reliable but absolute numbers are approximate.
• Billed benefit framed as “3× longer sessions” and lower costs.

Contradiction: Hook Can Raise Costs (confidence: high)

Issue #582 reports that RTK’s PreToolUse hook for Claude Code increased costs by 18% on an identical debugging task. The hook silently rewrites commands (cat → rtk read) via "permissionDecision": "allow", and the compressed output forces Claude to generate ~50% more output tokens to compensate for missing information. Net result: input savings < output increase. This contradicts the “cost reduction” headline and is the canonical example of the “token reduction ≠ cost reduction” anti-pattern.

Independent measurement (JetBrains series, Part 2, reported 2026-08-03): the same paired SkillsBench harness that measured ponytail (Part 3) put RTK at +7.6% cost across 80 paired tasks versus an advertised −60–90% — “every bit as significant” as ponytail’s saving, “just pointing the wrong way.” It corroborates issue #582 from a completely independent direction: RTK’s input-side compression does not reduce the bill. The advertised range is not reproduced on a benchmark the vendor did not choose.

Implications: two independent failures (an 18% hook-cost report and a +7.6% paired benchmark) now sit against a −60–90% marketing claim. This makes RTK the strongest negative evidence in the vault for token-reduction-is-not-cost-reduction: whatever the token-level compression does, billed cost does not follow. The asymmetry (input saved at ⅕ rate, output rebilled at 5× on Opus-class models) is the mechanism in both cases.

Implications (confidence: medium)

Compression tools are not free. Any tool that reduces input tokens must account for the output side of the ledger, because output tokens are billed at a higher rate and degraded input can trigger verbose re-reading. The 5× input/output cost ratio on Opus-class models makes this asymmetry decisive. Measure total cost, not token counts, before adopting a proxy.

Open Questions

• Whether RTK has since fixed the hook economics (issue #582 is closed).
• Whether bytes / 4 estimation misleads users who assume exact token billing.
• Whether the +7.6% measured cost is sensitive to workload type (JetBrains’ SkillsBench skews to data/analysis tasks; no countervailing large over-build scenario was tested).

Related

token-usage-reduction — the concept this tool targets; records the contradiction headroom — competing compression proxy that also trims output tokens token-optimizer-mcp — MCP-based optimizer with caching + model routing claude-code-system-prompt — the agent RTK hooks into ponytail — same JetBrains series; the only arm that produced a solid cost saving

Sources

• raw/external/github-com-rtk-a2ac6abb.md
• raw/external/rtk-ai-app-index-1d38dd66.md
• raw/external/github-com-f7d4988e.md
• raw/external/blog-jetbrains-com-ponytail-skill-claude-tested-b56c8490.md