fx: Vercel's Tiny Open-Source Coding Agent

fx is a free, open-source coding agent from Vercel Labs, built to be as small and fast as possible. It’s a single native binary written in Zig — no runtime to install, no heavy IDE-in-a-terminal interface, just a CLI that feels closer to a Unix shell than a chat app. It started as an internal Vercel tool and was recently open sourced under the Apache-2.0 license. This guide covers installing fx and running your first coding task with it.

Quick Summary: fx is a 6-8MB native coding agent CLI from Vercel Labs. It’s free, open source, model-agnostic (works with any AI provider), and designed to cold-start in microseconds rather than seconds. No product telemetry, no forced cloud dependency.

Most AI coding agents today are built as Node.js or Python apps wrapped in a heavy terminal UI. fx takes the opposite approach: minimalism as a design principle, applied to the system prompt, the tool set, and the binary itself. The result is something you can install in seconds, embed inside other tools, or even run entirely inside a browser tab via WebAssembly.

fx vs. Other Terminal Coding Agents

Tool	Built With	Binary Size	Model Choice
fx	Zig (native)	~6-8 MB	Any provider, local or cloud
Claude Code CLI	Node.js runtime	Requires Node.js installed	Anthropic models
OpenCode	Various runtimes	Larger footprint	Multiple providers via gateway
Cursor CLI	Editor-integrated	Tied to Cursor app	Cursor’s model catalog

What You’ll Need

macOS, Linux, or Windows (via WSL)
A terminal
An API key from at least one model provider (or a local inference setup)

Step 1: Install fx

Step 1Open your terminal and run the install script below. This downloads the single native binary — no package manager, no runtime dependencies.

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

Warning: fx is explicitly labeled experimental by Vercel Labs. Expect frequent changes and occasional rough edges — this is a research-oriented tool, not a polished consumer product yet.

Step 2: Run Your First Task

Step 2From inside a project folder, launch fx and describe what you want in plain English — for example, “find the bug causing the login form to submit twice.” fx reads relevant files, makes a plan, and edits code directly, all inside your terminal.

Step 3: Get Structured Output

Step 3If you’re scripting or embedding fx into another tool, use the JSON output mode instead of the interactive CLI. This returns structured data instead of formatted terminal text, making it easy to pipe into other programs.

fx ask –json “summarize the changes in the last commit”

Step 4: Connect It to Your Editor

Step 4fx supports the Agent Client Protocol (ACP), which lets it connect directly to editors and other ACP-compatible clients rather than living only in a standalone terminal window.

fx acp

Step 5: Try It in the Browser

Step 5fx also ships a WebAssembly build, so you can try the whole agent inside your browser with zero install — useful for a quick test drive before committing to a local setup. Visit fx.sh/try to test it directly.

Step 6: Extend It with Skills, Plugins, and MCP

Step 6fx’s core is intentionally minimal, but you can extend it with reusable skills, plugins, or by connecting external tools through MCP (Model Context Protocol). You can also delegate independent pieces of work to subagents rather than handling everything in one session.

Tip: fx is designed to run fully hermetic — meaning with local inference and auto-updates turned off, it makes zero outbound network calls except to the model you’re actually using. That’s a meaningfully different privacy posture than most cloud-first coding agents.

The fx Workflow at a Glance

Install via curl → Run fx in a Project Folder → Describe the Task → Review & Apply Changes → (Optional) Connect via ACP to Your Editor → Extend with Skills / Plugins / MCP

Commands to Try

fx ask “explain what this function does”
fx ask –json “list all TODO comments in this repo”
fx acp
fx ask “write tests for the auth module”

Troubleshooting Common Issues

Install script fails or is blocked: Corporate networks sometimes block piping curl directly into bash. Download the install script first, review it, then run it locally instead of piping directly.

fx feels unstable or behaves unexpectedly: This is expected given its experimental status — check the GitHub issues page for known problems before assuming it’s something on your end.

No model responds: Confirm you’ve set a valid API key for at least one provider, since fx itself doesn’t include a bundled model — it’s model-agnostic by design and needs you to supply one.

ACP connection to your editor doesn’t work: Confirm your editor actually supports the Agent Client Protocol; not all editors do yet, since ACP is still a fairly new standard.

Frequently Asked Questions

Is fx free? Yes, the fx binary itself is completely free and open source under Apache-2.0. You still pay your model provider directly for AI usage, same as with most agent tools.

Do I need Node.js or Python installed? No. fx is a single native binary written in Zig, so there’s no runtime to install separately — that’s one of its main design goals.

Can I use fx with any AI model? Yes, fx is explicitly model and provider agnostic, and supports both local and cloud inference.

fx: Vercel’s Tiny Open-Source Coding Agent