When your coding agent is smaller than the models it calls

What it does

hax is a terminal-native coding agent written in C that connects to LLM providers—local llama.cpp servers, OpenAI, Anthropic, OpenRouter, and others—and streams answers with live tool output directly in your shell. It runs as a single native binary using only a few megabytes of RAM, leaving most memory for the models themselves. You can use it interactively or pipe one-off prompts through it like a standard Unix filter.

The interesting bit

Instead of hijacking your terminal with custom renderers, hax reflows Markdown and live output to fit the display while preserving native scrollback, redrawing only the current streaming line or input area. It also exposes the full request and response history in a transcript view and optional wire traces, so you can audit exactly what the model saw.

Key highlights

• Single C binary with minimal dependencies; starts instantly and stays light on memory.
• First-class local model support that auto-discovers llama.cpp servers without extra configuration.
• Broad provider coverage including OpenAI, Anthropic, Codex, OpenRouter, and custom endpoints.
• Well-behaved Unix citizen: uses XDG paths, plain-text config, clean stdout in one-shot mode, and composes via subprocesses rather than plugins.
• Deliberately omits MCP marketplaces, IDE panels, and plugin runtimes—see docs/philosophy.md for the design rationale.

Caveats

• FreeBSD and OpenBSD builds are source-only; no prebuilt binaries.
• Windows support requires WSL; there is no native Windows build.
• @file completion depends on fzf being installed, though it is optional.

Verdict

Reach for hax if you live in the terminal, run local models on modest hardware, or want a coding agent that behaves like standard Unix tooling rather than a web app. Look elsewhere if you need IDE panels, per-command permission prompts, or a plugin marketplace.