rtk-ai/rtk

What it does

RTK sits between an AI coding agent and the shell. When the agent runs a command such as git status or cargo test, RTK executes it, filters the output, and passes a shorter version back, so the agent reads a 3-line summary instead of 40 lines of raw text (README). It covers 100+ commands across Git, GitHub CLI, Cargo/Rust, JavaScript, Python, Go, Ruby, .NET, JVM, PHP, Docker, Kubernetes, AWS, Terraform, and file/search tools, with per-command filters that must cut at least 20% to ship (command coverage).

It is a single Rust binary aimed at developers running AI coding agents who want to reduce input tokens without changing their workflow. RTK advertises up to 60–90% less command output, but the compressor only affects command output, one part of the input tokens and one part of the bill. Token counts are estimated as bytes/4, so percentages are directional and absolute numbers approximate (savings explanation).

Getting started

Install with Homebrew on macOS/Linux, winget on Windows, or the install script on Linux, macOS and WSL:

brew install rtk
winget install rtk-ai.rtk
curl -fsSL https://raw.githubusercontent.com/rtk-ai/rtk/refs/heads/master/install.sh | sh

Then register the hook for your agent and restart it:

rtk init -g          # Claude Code; other agents have their own flags

After that, commands the agent runs are rewritten to rtk <command>, and output is compressed automatically. Check savings with rtk gain, or rtk gain --graph for a 30-day ASCII chart; rtk bypass <command> runs a command unfiltered while still tracking it (README install section). Note there is another rtk on crates.io: rtk gain should show a savings dashboard, not "command not found". Agent-specific setup for Copilot, Cursor, Gemini CLI, Codex and others is in the supported agents guide.

Usage constraints

• The automatic hook is installed per agent; with rules-file agents RTK asks the model to prefer rtk commands rather than enforcing them (agent coverage).
• Built-in tools such as Read, Grep and Glob pass through untouched by default; compressing them requires a pipeline (README).
• Pipelines can rewrite commands and send content to the models they declare, so read a shared pipeline file before installing it (pipelines overview).
• The optional rtk-proxy daemon handles provider API traffic, is a separate binary and release archive, and is licensed BUSL-1.1 (proxy docs).
• Two rtk editions install the same binary name and cannot coexist: rtk bundles the pipeline engine under Apache-2.0 + BUSL-1.1, while rtk-oss is Apache-2.0 only (editions).
• Telemetry is anonymous and off by default; no command content or file contents are collected (telemetry).

HotRepo's take

The project's HotRepo record shows 50,542 stars on 2026-05-19 and 51,504 on 2026-05-20, gaining 667 and 704 stars on those daily all-language snapshots; by a 2026-11-13 Rust-language daily snapshot it had 86,426 stars and 5,540 forks, ranking 9th on that language list that day. Those points sit months apart, so they describe a long-run climb rather than a single trend window.

HN discussion adds a concrete engineering question: one commenter asked whether compressing CLI output could confuse models that were mostly trained on raw CLI output, since the agent may lose the exact format it knows (Oras, 2026-04-10). The README's own answer is that filters are written against real output and each cut must save at least 20%, but that is the maintainer's design claim, not a measured test of the training-data concern. A separate naming complaint says the existing RTK project on crates.io creates confusion (adithyassekhar, lsaferite). The repo has a rtk gain analytics path to verify savings on your own commands, which is the practical way to check whether the filtering works for your stack.