fx: a 6MB Zig coding agent built for embedding, not terminal theatrics

fx is an experimental, open-source (Apache-2.0) coding agent written in Zig that trades the heavy “IDE in the terminal” approach for something closer to a Unix shell. The whole harness ships as a single ~6.39MiB binary, cold-starts in microseconds, and holds a single-digit-megabyte memory baseline — design choices aimed at packing many instances onto one machine and dropping the agent into sandboxes, CI, or larger automated systems. The CLI leans on minimal output and preserved scroll history rather than full-screen TUI painting.

The project’s real pitch is embeddability and efficiency. A deliberately lean system prompt and small tool set cut token costs and improve time-to-first-token, while a Wasm build target (fx.wasm) shrinks the footprint further and makes the network stack pluggable. Extensibility follows a Unix philosophy: a small core extended through skills, plugins, and MCPs. It is model- and provider-agnostic, working against local models, gateways, or hosted APIs, which suits both offline and cloud inference.

At v0.0.4 the authors flag it as experimental and warn of frequent breaking changes, so it is more a research and integration substrate than a daily driver today. Still, the emphasis on tiny binaries, fast startup, and low resource use points at a niche the current crop of agent tooling largely ignores — running coding agents cheaply, programmatically, and at scale rather than interactively.