Vercel Labs open-sources fx: a 6MB native coding agent in Zig
A 6MB tool built to embed in bigger systems, still v0.0.4 — plus Unsloth's sharper local quants and OneCLI's Rust sandbox for team agents.
Copy markdown
Fast, native, and built to embed
fx is a CLI and harness written in Zig with ~10µs cold starts and a 6.39MB binary, built to be embedded in larger systems rather than run as a heavy TUI. It keeps a minimal system prompt and toolset to cut token costs, with a shell-like interface that preserves scroll history.
Model-agnostic by design, gateway-first today
Right now fx routes through the Vercel AI Gateway, with subscription support for Codex and Grok flagged as incoming; the architecture targets local models, gateways, and direct provider APIs. It's extensible via skills, plugins, and MCPs.
Install it and drop it in your stack
`curl -fsSL https://fx.sh/setup.sh | bash` gets you running; the source is Apache-2.0 on GitHub and compiles to native binaries and WebAssembly, so you can embed it in your own multi-agent setup.
The catch: v0.0.4 and Vercel gravity
fx is explicitly experimental — v0.0.4, with frequent breaking changes expected — and HN pushed back on the "tiny" claim (agents exist at 0.6MB) and the 26 bundled tools. Login currently expects a Vercel account.
Elsewhere: Unsloth Dynamic 3.0 sharpens local runs
Unsloth's new Dynamic 3.0 GGUFs claim up to +10% top-1% accuracy at the same file size versus rival quants, letting you run models like Qwen3.8-27B locally near 10GB with better fidelity. llama.cpp-compatible, no quant-aware training.
Elsewhere: OneCLI ships a sandboxed team agent harness
A Launch HN under Apache-2.0 and written in Rust, OneCLI gives each teammate an isolated-VM agent with credentials injected per-request at a gateway and policies enforced outside the model. Self-host or cloud.