Hax – a minimalist, terminal-native coding agent written in C | AI Paper Digest

TL;DR Highlight

A tiny single-binary C coding agent for the terminal that works out of the box with local models

Who Should Read

Terminal-focused developers who want to run a coding agent on local models (llama.cpp) without a heavy tool

Core Mechanics

Single C binary, very light
- It is one native C binary with few dependencies.
- It starts instantly and uses only a few MB of memory.
- That leaves more RAM for local models.
Local models are first-class
- It auto-discovers the model and runtime capabilities.
- The default setup needs no custom provider config block.
- Run llama-server, then hax --provider llama.cpp.
It doesn't mess with your terminal
- Streaming Markdown and tool output are reflowed for the terminal.
- Only the streaming line or input area is redrawn, so native scrollback is kept.
You can inspect what was sent
- Ctrl+T shows what was sent to the model and what it replied.
- A detailed wire protocol trace is optional.

Community Reactions

Used with deepseek-flash via the Anthropic-compatible provider
- A commenter built a Rust project they had put off for months.
- Cost was 1 US cent and setup took under 10 minutes.
Tried on macOS with llama.cpp
- A commenter built and installed it and found it clean and fast with local models.
Opinion on model bias toward their own agents
- A commenter says models are finetuned around their vendors' agents.
- Opus reportedly wants specific bash tools and writes Python calling bash if they are hidden.
Skepticism about C
- One commenter questioned starting a new project in C.
- Another confirmed interrupt handling lives in interrupt.c.

How to Apply

Hook hax up to llama.cpp
- Install with brew install oleksandrchekhovskyi/hax/hax.
- Start llama-server -m model.gguf and run hax --provider llama.cpp.
- Ask for a small edit in a project folder.
Use one-shot mode in scripts
- Run hax -p "list TODOs" to get only the final answer on stdout.
- Pipe it into other tools or cron jobs.
Debug prompts with Ctrl+T
- When the agent misbehaves, open the transcript.
- Check what was actually sent and fix the system prompt or tool output.

Try It

copy & run

# Paste into a terminal (macOS/Linux, with llama.cpp)

brew install oleksandrchekhovskyi/hax/hax

# In another terminal, start the local model server
llama-server -m {{model}}.gguf

# Run in your project folder
cd {{project}}
hax --provider llama.cpp -p "list TODOs"

# Next: run hax interactively and press Ctrl+T to view the transcript

Terminology

harnessThe wrapper program that gives an LLM tools and a loop so it works as an agent. Think of it as the chassis and cockpit.

llama.cppA C/C++ inference engine that runs GGUF models locally. llama-server exposes an API.

wire protocol traceA raw log of the requests and responses exchanged with the model API.

one-shot modeRuns a single prompt and prints only the final answer, which suits shell scripts.

Hax – a minimalist, terminal-native coding agent written in C