干净的上下文。
更强的代理。

RTK is a cost reduction engine for AI coding agents. It cleans what your agent reads at every tool call — shell output, file reads, search results, web pages, MCP calls — cutting up to 90% of command output noise, and it can hand heavy work to cheaper models. The battle-tested Bash compressor developers starred is the core of the engine. RTK is free to use, written in Rust, with the CLI open source under Apache 2.0, and works with Claude Code, Cursor, and 18 AI coding agents.

RTK Engine turns the battle-tested compressor into a full cost reduction engine. RTK can now compress every tool's output, not just shell commands, route heavy work to cheaper models, and let you build your own pipeline, plug in new methods, and share your setup as one file that works in any supported agent. Upgrading changes nothing: RTK works exactly as before until you add a line.

Read the RTK Engine announcement →

One file of rules. Each rule reads like a sentence: where RTK acts, when, and what it does — for example, when the agent reads a file longer than 350 lines, send it to a cheaper model. RTK's own behaviour is written in the same format, so you never start from a blank page: pick a built-in pipeline, change one line, or write your own.

Yes. A pipeline is one file: share it, drop it in, and it works in any supported agent. The same file can be used in Claude Code by one person and Cursor by another.

No. Install, run rtk init, and RTK works exactly as before. Nothing changes until you add a line to a pipeline. The compressor you already rely on is the core of the engine, not a feature that moved.

On the commands RTK rewrites, output drops by 56% on average, and by up to 90% on the noisiest ones. Across the community that adds up to billions of tokens saved. The exact figure depends on the command, so the benchmarks page publishes the methodology and the per-command results, and the blog covers what independent studies measured, including the cases where RTK did not come out ahead. For your own numbers, run rtk gain after a few days of use.

See the benchmarks →Read the analysis →

RTK is completely free. It is open source under the Apache 2.0 license, with source code available on GitHub at github.com/rtk-ai/rtk. There are no usage limits, no API keys required and no accounts. Routing to a cheaper model is optional and uses a model key you provide. Telemetry is opt-in: it is off until you explicitly enable it, and it collects no personal data. For teams, RTK Pro is available at pro.rtk-ai.app.

RTK works with 18 AI coding agents, including Claude Code (Anthropic), Cursor, GitHub Copilot, Codex, Gemini CLI, Windsurf, Cline and Aider — and the engine acts on every tool an agent uses: shell commands, file reads, search, web fetches and MCP calls. On the shell side it covers 125+ commands: cargo test, pytest, go test, git diff, git status, git log, grep, find, ls, pnpm list, tsc, eslint, prisma, docker, kubectl, and more.

Running rtk init --global installs hooks inside your agent that act around each tool call: shell commands are automatically rewritten to rtk equivalents, and output is compressed before it enters the context window. If you turn it on, the RTK Proxy also acts on the API traffic, trimming what the agent sends to the model. You continue using your normal commands, with no manual changes to prompts or workflows needed.

No. RTK removes verbose boilerplate and repetitive output noise, not meaningful content. Test failures, error messages, diffs, and stack traces are preserved in full. The AI receives the same essential information with far less noise around it, which leaves more of the context window for reasoning.