tokless

README

A unified pipeline for efficient and effective coding agents.

One tool, no config — works the moment it lands.

Introduction

Many great packages make coding agents more effective and efficient — but discovering, installing, updating, and unifying them is painful, especially for non-technical users. The best tools exist; the wiring is the real cost.

tokless — the lazy one-command solution.

✔️	Best packages, unified — picks the most effective, efficient tools and wires them without conflicts
✔️	One command, done — pick your agent, restart, go
✔️	All platforms — macOS, Linux, Windows
✔️	Zero config — everything wired, no manual edits
✔️	Simple updates — tokless update upgrades everything in one shot
✔️	Non-tech friendly — under 30 seconds, anyone can do it

Installation

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/HoangP8/tokless/main/scripts/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/HoangP8/tokless/main/scripts/install.ps1 | iex

Supported Agents

Claude Code ✓ Done	OpenCode ✓ Done	Codex ✓ Done	Antigravity ✓ Done
Pi ✓ Done	Cursor In progress	Factory Droid ✓ Done	GitHub Copilot ✓ Done

Pick one, some, or all:

tokless                              # interactive: pick agents
tokless --agents claude,opencode     # wire just these
tokless --agents claude,opencode,codex,antigravity,copilot,droid,pi  # all

Tools

Tool	⭐	What it does
karpathy-skills		Distilled meta-rules from Karpathy's LLM-coding post — think before coding, simplicity first, surgical changes, goal-driven. Drops overbuild and wrong-assumption failures.
caveman		Skill/plugin forcing terse caveman-speak across 30+ agents — 65% output token cut, technical content untouched.
ponytail		Skill embedding a lazy senior dev — minimum-code, stdlib-first, no speculative features across 16 agents.
rtk		CLI proxy filtering/compressing command output before it hits the LLM; 100+ commands, single Rust binary, <10ms overhead.
codegraph		Pre-indexed SQLite code knowledge graph — auto-syncs on file change, returns verbatim source + call paths via 1 MCP tool across 8+ agents.
context-mode		MCP server for 17 clients — sandboxes tool output, persists session memory, enforces think-in-code.

Configuration

Each tool is wired into each agent through the agent's native config system:

Tool	Claude	OpenCode	Codex	Antigravity	Copilot	Droid	Pi
rtk	Hook + Allow	Plugin	Hook	Hook + Allow	Hook + Allow	Hook	Extension
caveman	Plugin + Instruction	Plugin + Instruction	Skill + Instruction	Skill + Instruction	Skill + Instruction	Skill + Instruction	Skill + Instruction
ponytail	Plugin + Instruction	Plugin + Instruction	Plugin + Instruction	Plugin + Instruction	Skill + Instruction	Skill + Instruction	Skill + Instruction
codegraph	MCP + Allow + Instruction	MCP + Instruction	MCP + Instruction	Hook + MCP + Instruction	Hook + MCP + Instruction	Hook + MCP + Instruction	MCP + Extension + Instruction
context-mode	MCP + Allow + Instruction	Plugin + Instruction	Hook + MCP + Instruction	MCP + Instruction	MCP + Hook + Instruction	MCP + Instruction	MCP + Instruction

Usage

tokless              Install + wire everything (default; safe to re-run)
tokless update       Update the tokless CLI, then show version diff and upgrade tools
tokless doctor       Show what's wired; warn about broken bits
tokless info         Show how tokless was installed, plus paths and config locations
tokless index        Build per-project codegraph indexes
tokless disable      Disable one or more agents
tokless uninstall    Remove everything tokless touched
tokless self-update  Update the tokless CLI itself
tokless --version    Print tokless version
tokless --help       Show all commands and flags

Flags:

--agents <list>   Subset: claude,opencode,codex,antigravity,copilot,droid,pi
--tools <list>    Subset: rtk,caveman,ponytail,codegraph,context-mode
--dry-run         Preview, no writes
--verbose         Every step
--yes             Skip confirmations

Restart agents after install so they pick up new config.