Hax: A Lean, C-Powered Terminal Coding Agent

hax is a minimalist, terminal-native coding agent engineered entirely in C. It is designed for those who prefer a stripped-down, high-performance interface over bloated IDE integrations.

"If 'fancy new AI tech in an old-school minimalist package' sounds like your vibe, you might like this."

---

🚀 Core Philosophy & Design

Hax is built to be invisible yet powerful. Unlike modern agents that bundle heavy runtimes, Hax focuses on efficiency and the Unix philosophy.

Lightweight Architecture

By utilizing a single native C binary with minimal dependencies, Hax ensures that system resources are preserved for the actual intelligence—the LLM.

$\text{Available RAM} = \text{Total RAM} - (\text{LLM Memory} + \text{hax Memory})$

Since hax only consumes a few MBs, the $\text{hax Memory}$ variable is negligible, leaving maximum headroom for local models.

The "Minimalist" Approach

Hax deliberately avoids "feature creep." Here is how it compares to traditional AI agents:

What Hax explicitly excludes:

• MCP Marketplaces
• Plugin Runtimes
• IDE Side-panels
• Per-command permission prompts

Note: The reasoning for these omissions is detailed in the docs/philosophy.md file.

---

🛠 Technical Features

Local Model Integration

Local LLMs are treated as first-class citizens. Hax automatically detects runtime capabilities, meaning you don't need complex provider configuration blocks for standard setups.

Example Workflow:

1. Start your server: llama-server -m [model].gguf
2. Launch the agent: hax --provider llama.cpp

Terminal-First UX

Hax respects the terminal environment. It provides Streaming Markdown and live tool output that is reflowed to fit your window. To maintain a clean history, it only redraws the active streaming line or the input area, ensuring your native scrollback remains intact.

Transparency & Compatibility

You can audit every interaction. By pressing Ctrl+T, you enter a transcript view to see the exact prompts sent and the raw responses received.

Unix Compliance

Hax behaves like a proper Unix tool:

• Follows XDG paths.
• Uses plain-text for configuration and session files.
• In -p (one-shot) mode, stdout remains clean, while resume hints are sent to stderr.
• Prefers subprocess composition over a proprietary plugin system.

---

📦 Installation

Hax is available for Linux and macOS. Windows users should run it via WSL.

Option 1: Quick Install

• macOS/Linux (Homebrew): brew install oleksandrchekhovskyi/hax/hax
• Linux (Static Binary): Download the x86_64 or aarch64 binary from the latest release and move it to a directory in your PATH.

Option 2: Build from Source

Building from source links the binary against your specific system libraries.

# Clone the repository
git clone https://github.com/OleksandrChekhovskyi/hax.git
cd hax

# Install dependencies (C compiler, libcurl, jansson, meson, ninja, pkg-config, fzf)
scripts/install_deps.sh 

# Compile and install
make
make install # Optional; may require sudo

---

⌨️ Usage Guide

Run hax from within the project directory you wish to analyze.

Command Reference

First-Run Tips

The most intuitive way to start is the interactive mode:

1. Run hax.
2. Type /provider to view and select your preferred model/provider.
3. Hax will remember your provider, model, and effort settings for future sessions.

For deeper configuration, refer to the README and full documentation.