Slash Tokens

CLI proxy that reduces LLM token consumption by 60-90%. Prefix any dev command with 'rtk' to get filtered, compact output. Use for all Bash commands to save...

Security Scan

The tool appears to do what it claims (a local CLI proxy that filters command output), but there are several mismatches and risky instructions (remote install curl|bash, telemetry/analytics, auto-rewrite hooks, and repo/URL inconsistencies) that warrant caution before installing.

Purpose & Capability

The name/description (CLI proxy to reduce LLM tokens) matches the included source: many command filter modules, tracking/analytics, and a hook system. HOWEVER there are incoherences: the SKILL.md and registry say 'instruction-only / no install spec' while the package contains 100+ source files (Rust, scripts, an openclaw plugin). The SKILL.md's recommended install curl URL points at raw.githubusercontent.com/2233admin/rtk whereas Cargo.toml points to github.com/rtk-ai/rtk — this repository/URL mismatch is unexplained and suspicious.

Instruction Scope

SKILL.md tells users to run curl -fsSL https://raw.githubusercontent.com/2233admin/rtk/main/install.sh | bash (pipe-to-shell). The runtime docs also describe auto-rewrite hooks that can silently intercept and rewrite commands (Auto-Rewrite default). The skill documents writing/reading a SQLite DB at ~/.local/share/rtk/history.db and includes commands to export and publish those stats. While filtering output is expected, the instructions encourage executing an external install script and enable behavior (automatic command rewriting, global history collection) that can affect many commands and leak local outputs if telemetry/export is enabled.

Install Mechanism

No formal install spec is present in the registry entry but SKILL.md instructs users to curl|bash an install.sh hosted on raw.githubusercontent.com under a different username (2233admin). Running an arbitrary raw GitHub script is high-risk. The repo also contains an install.sh, Homebrew formula, and packaging metadata — but the mismatch in authoritative repository/URL should be resolved before trusting an automated install.

Credentials

The skill declares no required environment variables or credentials, which is consistent for a local CLI proxy. However source files and docs show persistent tracking (SQLite history at ~/.local/share/rtk/history.db), telemetry modules (src/core/telemetry.rs) and analytics (src/analytics/*) with HTTP dependency (ureq). That implies device identifiers/hashed salts and command outputs could be collected and (potentially) transmitted; the skill does not request credentials but it does collect command output and metadata globally across projects.

Persistence & Privilege

always:false (good), but the code and docs show persistent local state (history.db, salt files with 0600 perms) and a hook system that can auto-rewrite commands before execution. Auto-Rewrite is documented as the default with '100% adoption' (silent interception). Silent rewriting of user commands and persistent tracking raise the blast radius if telemetry/networking or malicious install sources are involved.

What to consider before installing

Do not pipe the remote install script to bash without verification. Before installing: 1) Verify the canonical repository and install URL (Cargo.toml repository vs SKILL.md curl URL mismatch). 2) Inspect install.sh in that authoritative repo (or build from source) instead of running it directly. 3) Check telemetry & analytics code (src/core/telemetry.rs, scripts) and determine whether you can opt out; verify what endpoints it calls. 4) Be aware rtk stores a global history DB at ~/.local/share/rtk/history.db (contains raw outputs/estimates) and has commands to export/publish that data — treat it as potentially sensitive. 5) Note the hook auto-rewrite behavior: if you need guarantees the original command runs, use verbose/proxy modes or disable hooks; confirm defaults before using in CI or on machines with sensitive commands. If you want to proceed, prefer building/installing from a verified source (git clone the trusted repo and cargo build / inspect scripts) and audit the install.sh and telemetry endpoints first.

Slash Tokens — 你的 AI 编程账单砍九成

故事

有个哥们用 Claude Code 写代码，一个月烧了 $200。他不理解——明明就改了几个文件，钱去哪了？

他翻了翻对话记录，发现了一个惊人的事实：80% 的 token 不是他写的代码，是工具输出。

git log 吐了 200 行，Claude 只需要看最后 5 行。npm install 输出了 800 行进度条，Claude 一行都不需要。docker ps 带了一堆格式化边框，全是废话。

这些垃圾输出塞满了 context window，Claude 被迫"阅读"每一行，然后 Anthropic 按 token 收费。

就好比你请了个律师，按小时收费，然后你让他坐在那儿读你的垃圾邮件。

于是他写了 rtk。

一个 Rust 写的 CLI 代理，4MB，零依赖。在命令前面加个 rtk，输出就被过滤成 AI 真正需要看的部分。

git status 吐 30 行？rtk git status 给你 3 行。 npm test 输出 500 行？rtk test 只给你失败的那 2 个。 kubectl get pods 带一堆表格边框？rtk kubectl get pods 干干净净。

结果：月账单从 $200 掉到 $40。省了 80%。

那个哥们后来把 rtk 开源了。现在支持 61 个命令，覆盖你日常用的所有工具。

Install

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/2233admin/rtk/main/install.sh | bash

# Windows (scoop)
scoop install rtk

# ClawHub (skill only, binary installed separately)
clawhub install slash-tokens

Core Usage

在命令前加 rtk，完事：

rtk git status          # 30 行 → 3 行
rtk git log --oneline   # 去掉装饰符
rtk gh pr list          # token 友好的 PR 列表
rtk npm test            # 只看失败
rtk docker ps           # 紧凑容器列表
rtk pytest              # 只看红的
rtk kubectl get pods    # 去掉表格边框

Key Commands

Command	What it does
rtk git	Git with compact output
rtk gh	GitHub CLI, token-optimized
rtk test / rtk pytest / rtk vitest	Tests — only show failures
rtk err <cmd>	Run anything, show only errors/warnings
rtk diff	Ultra-condensed diff (changed lines only)
rtk json <file>	Compact JSON, --schema for structure only
rtk find	Compact file tree
rtk grep	Strips whitespace, truncates, groups by file
rtk curl	Auto-JSON detection + schema output
rtk summary <cmd>	Run anything, get 2-line heuristic summary
rtk gain	Show token savings stats
rtk discover	Find missed savings from past sessions

Supported Tools (61 commands)

git, gh, docker, kubectl, cargo, npm, npx, pnpm, pip, go, pytest, vitest, playwright, ruff, mypy, tsc, next, lint, prettier, prisma, psql, aws, curl, dotnet, rake, rubocop, rspec, golangci-lint, and more.

When to use

Prefix ALL Bash commands with rtk to save 60-90% tokens. If rtk doesn't have a specialized filter for a command, use rtk err (errors only) or rtk summary (2-line summary) as fallbacks.