Vercel fx

Canonical version: Vercel fx.

fx is a coding agent CLI from Vercel Labs, written in Zig. The pitch fits in one line: "Tiny, open, embeddable, native coding agent." The whole thing is a ~6 MB binary, Apache-2.0 licensed, and it runs on macOS and Linux. It's still experimental (v0.0.x at the time of writing), and Vercel says it plainly: use at your own risk, expect frequent changes.

Install: curl -fsSL https://fx.sh/setup.sh | bash. Then fx in a project folder for the interactive shell, or fx ask "..." for one-shot requests in scripts.

Not to be confused with Anton Medvedev's fx, the terminal JSON viewer (fx.wtf).

A shell, not an IDE

Most coding agents want to be your IDE inside the terminal. fx goes the other way: it behaves like a Unix shell. Output renders inline and your scrollback stays intact. That sounds minor, but if you live in the terminal, it really is a big deal.

The more interesting angle is embeddability. fx is designed as an AI Agent Harness that you drop into larger systems, not just a tool you type into:

  • fx acp connects it to editors and any Agent Client Protocol (ACP) client
  • createFxAgent() embeds the agent core in a JavaScript host via fx-core.wasm
  • createFxTerminal() embeds the full interactive terminal via fx-term.wasm
  • The SDK ships on npm as libfx, with Node.js, browser, Next.js and Nuxt examples

The host application can provide its own network transport, session storage, configuration, permission handling and terminal I/O. The fx.sh homepage demo actually runs the complete CLI as WebAssembly in your browser, with networking delegated to fetch. That's a nice proof of the design.

Models and extensions

fx is model-agnostic. You can sign in with the Vercel AI Gateway (fx login), a ChatGPT subscription through Codex OAuth (fx login codex), a Grok subscription (fx login grok), or point it at any OpenAI-compatible endpoint like Ollama or OpenRouter via named connections in ~/.fx/settings.json. When going through the AI Gateway, you can even set provider routing order (e.g., try Bedrock first, then Anthropic).

It supports the usual extension points: skills, MCP servers and subagents. There's also a Slack bot integration.

Small detail I appreciate: every docs page is available as Markdown by appending .md to the URL, and there's an llms-full.txt with everything. Docs written for agents, not just humans.

My take

fx competes less with Claude Code than with harnesses like Pi Mono: small, hackable cores meant to be composed into your own tooling. The fact that the same binary targets native and WebAssembly opens doors (agents running in the browser, in sandboxes, inside other apps). Too early to bet on it, but worth watching.

References


About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe