Miskin

Save AI tokens across providers. Local. No server. Compresses command output + injects caveman-mode prompts.

Without Miskin:                     With Miskin:

git status ──▶ shell ──▶ LLM        git status ──▶ miskin ──▶ git ──▶ filter ──▶ LLM
  (~2000 tokens)                      (~200 tokens, 90% saved)

Inspired by RTK and Caveman.

Quick Start

# Install
npm install -g miskin                    # npm
# or: cargo install miskin               # crates.io

# Install hooks for your AI tool
miskin init                # Claude Code (default)
miskin init --agent cursor
miskin init --agent copilot
miskin init --agent gemini
miskin init --agent codex
miskin init --agent opencode

# Restart your AI tool (some load hooks automatically), then test
miskin git status          # 90% fewer tokens than raw git status

How It Works

Two strategies, one binary:

Strategy	What	Token savings
Command filter	Run command, compress output before LLM sees it	60-90% on dev commands
Caveman prompt	Inject "be concise" into system prompt	65% on LLM responses

The hook rewrites bash commands transparently:

git status    →  miskin git status
cargo test    →  miskin cargo test
ls -la        →  miskin ls -la
docker ps     →  miskin docker ps

Miskin runs the real command, filters the output through format-specific compressors, and returns compact results. The LLM gets the same information in fewer tokens.

Commands

# Passthrough mode — run any command through a filter
miskin git status
miskin git diff
miskin git log --oneline -10
miskin cargo test
miskin cargo build
miskin cargo clippy
miskin npm install
miskin docker ps
miskin ls -la
miskin cat src/main.rs
miskin pytest
miskin jest
miskin go test
miskin eslint .
miskin ruff check .
miskin curl https://api.example.com

# Subcommands
miskin init              # Install hooks
miskin init --show       # Verify installation
miskin stats             # Token savings report
miskin gain              # Alias for stats
miskin stats --daily     # Day-by-day breakdown
miskin stats --graph     # ASCII bar chart
miskin stats --json      # JSON export
miskin discover          # Find unfiltered commands
miskin session           # Adoption across sessions
miskin compress          # Print caveman system prompt
miskin config show       # View configuration
miskin config set caveman.level aggressive
miskin proxy <cmd>       # Raw passthrough + tracking
miskin err <cmd>         # Show errors only
miskin completions bash  # Generate shell completions

Installation

npm (recommended)

npm install -g miskin

Cargo

cargo install miskin
# or from source:
cargo install --git https://github.com/sonyarianto/miskin

From source

git clone https://github.com/sonyarianto/miskin
cd miskin
cargo install --path .

Shell completions

# Bash
echo 'eval "$(miskin completions bash)"' >> ~/.bashrc

# Zsh
echo 'eval "$(miskin completions zsh)"' >> ~/.zshrc

# Fish
miskin completions fish > ~/.config/fish/completions/miskin.fish

Supported Commands (40+)

Category	Commands
Git	status, diff, log, add, commit, push, pull, branch, checkout, merge, rebase, stash, tag, fetch, clone, init, remote, reset, restore, rm, mv, clean
Rust	cargo build, test, clippy, fmt, check, run
Node.js	npm, pnpm, yarn, npx, bun, pip, uv, bundle, gem
Testing	pytest, jest, vitest, go test, rspec, playwright
Linting	eslint, ruff, biome, tsc, mypy, prettier, golangci-lint, rubocop, clippy
Docker/K8s	docker ps, images, logs, compose; kubectl, oc
Files	ls, cat, find, tree, head, tail, read
GitHub CLI	gh pr list/view/status, issue list, run list, repo view, auth
Network	curl, wget
System	df, du, ps, top, wc, env, which, uname, free

Approximate Savings

Note: These are rough projections based on typical output sizes, not measured benchmarks. Actual savings depend on project size, command flags, and output length. Run miskin stats to see your real numbers.

Command	Typical raw	With Miskin	Approx. saved
git status	up to ~2,000	~200	~90%
git diff	up to ~1,000	~300	~70%
git push	~200	~20	~90%
cargo test	up to ~5,000	~500	~90%
ls -la	~400	~100	~75%
docker ps	~500	~100	~80%
eslint	up to ~2,000	~400	~80%
npm install	~1,000	~50	~95%

Run miskin stats to see your actual numbers.

Caveman Modes

Inject into system prompts to make the LLM produce fewer tokens:

Level	Description
lite	Concise. No pleasantries.
full	Caveman speak. Fragments. No meta. (default)
ultra	Absolute minimum. Drop articles. One-word when possible.
aggressive	Ultra + code-body stripping in explanations.

miskin config set caveman.enabled true
miskin config set caveman.level ultra
miskin compress              # Preview the system prompt

Configuration

~/.config/miskin/config.toml:

[general]
enabled = true
ultra_compact = false
exclude_commands = ["curl"]  # Skip filtering for these

[filters]
max_lines = 500
deduplicate = true

[caveman]
enabled = false
level = "full"

[analytics]
enabled = true
data_dir = "/home/user/.local/share/miskin"

Analytics

Miskin tracks anonymized command token counts locally. No telemetry. No network calls.

miskin stats              # Summary
miskin stats --graph      # Per-command bar chart
miskin stats --daily      # Day-by-day
miskin stats --json       # JSON export
miskin discover           # Commands not yet filtered
miskin session            # Adoption by session

Data stored in ~/.local/share/miskin/analytics.json.

Supported AI Tools

Tool	Command
Claude Code	miskin init
GitHub Copilot	miskin init --agent copilot
Cursor	miskin init --agent cursor
Gemini CLI	miskin init --agent gemini
Codex	miskin init --agent codex
OpenCode	miskin init --agent opencode
Windsurf	miskin init --agent windsurf

Uninstall

miskin init --uninstall           # Remove hooks
miskin init --agent cursor --uninstall
cargo uninstall miskin

License

MIT