Open sourceRust · ≈85k stars · Apache-2.0
RTK
Rust CLI proxy that compresses shell-command output before it reaches a coding agent's context.
RTK is an agent-agnostic token-optimisation proxy shipped as a single Rust binary with no dependencies. It intercepts shell commands and compresses their output before the agent reads it. It supports 100+ commands with under 10 ms of overhead, and its GitHub description claims a 60-90% cut in token use on common dev commands. Examples include tree-format ls, signature-first file reads, grouped grep matches, compact git status, diff and log, and failures-only output for cargo, npm, pytest and go test. The README notes that savings apply to bash output only and are estimated as bytes/4, not a full-bill reduction.
Pros & cons
Each point cites its sources and says what kind of evidence backs it. A second reviewer tried to refute every claim before it was kept.
Pros
- DocsA single Rust binary with a hook that transparently rewrites shell commands for 18 agents (Claude Code, Codex, Cursor, Gemini CLI and others).[1]
- User reportsUsers find it genuinely useful for compressing repetitive output such as test runs on whitelisted commands.[1][2]
- DocsTelemetry is off by default and never collects command or file contents.[1]
- DocsThe README says plainly that savings cover only bash output and are estimated as bytes/4.[1]
Cons
- Independent testJetBrains' paired benchmark (425 billed trials) found no saving: median cost per task rose 7.6% at low effort (p=0.004) and was flat at high effort.[1][2]
- Independent testThe "rtk gain" counter overstates savings because it counts full raw output as the baseline: it reported 96.2M tokens saved in a run whose bill went up.[1][2]
- DocsThe hook skips Claude Code's built-in Read, Grep and Glob tools, and a user reports that persisting stats can break sandboxing and trigger auto-mode denials.[1][2]
- DocsMixed licensing: the CLI is Apache-2.0, but the compression engine and proxy are BUSL-1.1 source-available.[1]
- Best for
- Shell-heavy agent workflows (test runners, builds, git) where command output really dominates context.
- Watch out for
- Measure your actual billed cost with and without it, because the built-in savings counter is not a cost metric.
Assessed Oct 29, 2026 · adversarially reviewedSuggest a correction
People
Founder & CEO, rtk-ai
Cost & limits
Neighbours on the workbench
TypeScript · ≈393k stars · MIT
openclaw/openclaw
The viral open-source personal AI assistant that really does things.
Maintained by OpenClaw Foundation
MDX · ≈307k stars · MIT
obra/superpowers
An agentic skills framework and software methodology that works.
Maintained by Jesse Vincent (obra)
Shell · ≈306k stars · MIT
mattpocock/skills
Small, composable agent skills for real engineering work.
Maintained by Matt Pocock (AI Hero)
Connections
Relations from the Harness Hub knowledge graph, as charted in the State of Harness reports.
Maintained by · 1
- rtk-aiOrganisation
Sources
Verified Oct 29, 2026