nadezhdkov/claude-toolkit

**General-purpose tools, hooks, skills and subagents that make Claude Code more effective in any project.**
Why · Install · Commands · Roadmap · Design · Contributing
**Languages:** **English** · Português (Brasil) · Español · Deutsch · 简体中文
---
Why
AI coding agents lose time and make mistakes in the same places in almost
every project. claude-toolkit targets four of them, each measured on real
Claude Code sessions:
| Pain | Example | Tool |
|---|---|---|
| Builds and tests fail for environment reasons | Node 26 installed while the repo pins 24; node_modules missing | ctk doctor |
| Long command output floods the context window | Thousands of test lines to find three failures | ctk runq |
| Code navigation by grep instead of definitions | Dozens of file reads to find one pattern | ctk lsp |
| No way to tell whether a change helped | "It feels faster" | ctk sessions |
Every tool follows the same rules: it is **fast**, **quiet when everything is
fine**, and **never changes your system without asking**. Without a terminal
(for example when Claude runs it) a change prints its plan and exits with code
3, so a person approves it and reruns with --yes.
Output is **styled for people and plain for agents**: colours, charts and a
progress spinner appear only on an interactive terminal; pipes, logs and
Claude Code get compact plain text. NO_COLOR turns colour off and
FORCE_COLOR turns it on.
Install
Requirements: Linux, macOS or Windows (with Git Bash), git, and uv
(the installer offers to install uv if it is missing). The ctk command has no
runtime dependencies beyond Python 3.11+.
On Windows, in PowerShell: .\install.ps1 (same plan and questions; -Yes, -DryRun).
The installer runs uv tool install --editable . (so git pull updates
ctk) and then ctk setup, which adds a short section to
~/.claude/CLAUDE.md telling Claude Code when to use each command (with a
backup) and offers to install missing optional dependencies.
See docs/installation.md for manual steps, updating
and uninstalling.
Commands
Run ctk on its own for an overview:
| Command | What it does |
|---|---|
| ctk doctor | Toolchain pins vs installed versions, and dependencies vs lockfiles |
| ctk runq | Runs a command; prints failures, summary and the full-log path |
| ctk lsp | Language servers and Claude Code LSP plugins a project needs |
| ctk sessions | Counts-only metrics from Claude Code transcripts |
| ctk shoot | Deterministic screenshots of web pages, Godot scenes and desktop apps; baselines and contact sheets |
| ctk audiocheck | Loudness, clipping, silence and loop seams of audio files, with waveform and spectrogram images |
| ctk sandbox | Runs untrusted code without network, with a read-only system, an empty home and resource limits |
| ctk libdocs | Installed version, declared constraint, README and API of a dependency; where a symbol is defined |
| ctk map | Compact map of a repository's files and central symbols; where a symbol is used and which tests touch it |
| ctk verify | What is verified for the code as it is now: every runq run recorded per code state; fresh, failing, stale or never run |
| ctk diffcheck | Deterministic checks of a change: removed API still in use, risky files, debug leftovers, code without tests |
| ctk relay | Hands the work over to a fresh Claude Code session in a new terminal, from the handoff file and the verified state |
| ctk completion | Tab completion for bash, zsh, fish and PowerShell, generated from the commands themselves |
| ctk setup | Wires the toolkit into Claude Code, asking before each change |
| ctk hook | Logic behind the plugin hooks (event JSON on stdin) |