Vercel Labs  ·  Experimental  ·  Apache-2.0

One small binary.
A full coding agent.

fx is a tiny, open, embeddable coding agent harness and CLI — written in Zig, model-agnostic, and compiled to WebAssembly. Closer to a Unix shell than an IDE. And yes: it is running on this page.

curl -fsSL https://fx.sh/setup.sh | bash
Run fx in your browser vercel-labs/fx ↗
Binary
7.8 MiB
Startup budget
2 ms
Models
228 via Gateway
Runtime deps
0

The centerpiece

Don't read about it.
Run it.

Below is fx-term.wasm — the real interactive fx terminal, compiled from the same Zig source as the native CLI — booted inside your browser. No install, no server round-trip. Or take the guided session to watch a full agent loop play out.

simulated transcript · styled on real fx output

The WebAssembly build omits native processes, OS sandboxing and MCP by design — the host provides transport, storage and terminal I/O. SDK docs ↗

How fx works

An agent is a loop.

No framework, no runtime, no mystery. fx keeps the whole architecture small enough to hold in your head: prompt in, tools out, repeat until done.

  1. 01

    Prompt in

    You type intent in a composer that behaves like a shell — history, editing, slash commands. The workspace is simply your current directory.

  2. 02

    Model thinks

    One streaming request through Vercel AI Gateway — or a local model. fx is model-agnostic: FX_MODEL swaps the brain, not the harness.

  3. 03

    Tools out

    The model calls real tools — read, write, search, run. In auto mode, unresolved sensitive actions pause for your review before they execute.

  4. 04

    Result returns

    Tool output goes back into the context as plain data. The model sees what actually happened — exit codes, diffs, test failures.

  5. 05

    Repeat until done

    The loop turns until the task completes or hits the step limit. Every turn is recorded, resumable, and inspectable with /trace.

┃
auto · glm-5.2

7.8 MiB, all in

Smaller than the average photo.

The entire agent — parser, tools, permissions, terminal, MCP client, WASM target — ships as one dependency-free binary with a startup budget enforced in CI.

130× smaller than a typical Electron-based AI IDE — and fx is the agent.

Approximate download/install sizes for illustration. fx is measured; others are rounded 2026 figures.

Native or WebAssembly

One artifact, every surface.

Because the agent core compiles to a single .wasm module, fx drops into anything that can host WebAssembly or speak JSON-RPC. Hover a surface.

fx-core.wasm fx-term.wasm fx acp
In a browser host

Form factor

A Unix citizen, not an IDE.

Inline output in your scrollback. Composable subcommands. Plain JSON when you ask for it. These are real outputs from the real binary — click one.

Get started

Three commands. Done.

  1. 1

    Install

    curl -fsSL https://fx.sh/setup.sh | bash

    macOS and Linux, x86_64 and arm64. One static binary, no dependencies.

  2. 2

    Sign in

    fx login

    Sign in with Vercel, or run fx setup to drop in an AI Gateway API key.

  3. 3

    Run

    cd your_project && fx

    Your directory becomes the workspace. Type a prompt, browse /help, or one-shot with fx ask "…".

fx is experimental. It starts in auto permission mode, which reviews unresolved sensitive actions before they run. Use at your own risk, and tell us what breaks at fx.sh/feedback.

View more demos Get $10 off Kimi K3