fx — Vercel Labs' Unix-Like Coding Agent
Why This Matters
Most terminal agents take over the screen with a full UI.
fx from Vercel Labs behaves more like an ordinary command: the conversation prints into your terminal's scrollback, so you can scroll, search, and copy it, and tmux keeps working. It is one native Zig binary with no runtime to install, and scripts can call it like any other program. The formula pull request opened on August 22, 2026, at version 0.0.5 and merged early on August 23 UTC. The repo itself is only from August 11. Apache-2.0, 3,745 stars. The details below reflect 0.0.16, checked November 3, 2026.
What It Actually Does
- Scriptable.
fx ask --json "summarize the failing tests"writes the result to stdout and diagnostics to stderr, and exits nonzero on failure. - Embeddable.
fx acpserves editors over the Agent Client Protocol.libfxruns the same agent core from JavaScript, natively or in WebAssembly. - Your own shell. Commands run in your login shell with your aliases and
PATH. - Small on purpose. Per the README, CI fails a pull request if a benchmarked command such as
fx helptakes longer than 2 ms on Linux. The changelog puts the macOS arm64 binary at 5.97 MiB. - Shared conventions. It follows
AGENTS.md, including nested files, and discovers skills from.claude/skills,.agents/skills, and their home-directory equivalents.
The 0.0.16 changelog claims 41% fewer uncached input tokens than 0.0.15 across its caching scenarios. That is Vercel's own comparison between its own versions.
Caveats
The README's status line is "Experimental. Use at your own risk," and a 0.0.x version number agrees. GitHub's contributor list shows one account,
fazxes , with 3,753 commits, against 82 for the next, so this is close to a one-person project under the Vercel Labs name. The default permission mode deserves a look. In
auto , routine work such as editing project files and running tests goes ahead without asking. Anything else is first checked by a model for prompt injection or malicious behavior. Switch to ask if you want to approve each action. Sign-in is the other thing to check. The documented options are Vercel AI Gateway, a ChatGPT or Grok subscription, or an OpenAI-compatible endpoint such as Ollama or OpenRouter. There's no Claude subscription login. The docs do show Claude models by ID through AI Gateway, so Claude here means gateway billing rather than an existing plan.
The formula also conflicts with Homebrew's existing
fx formula, because both install an fx binary. Where It Fits in Our Workflow
We haven't run it. It wouldn't replace Claude Code for us, partly because of that billing point and partly because at 0.0.16 it's too early to build on.
It looks unusually easy to trial, though. Our
AGENTS.md should work as it is. Our ~/.claude/skills entries are symlinks into ~/code/lossless-monorepo/context-v/agent-skills/ , and since fx allows links that resolve inside your home folder, the docs suggest they'd load without extra configuration. That would make it a cheap test of how portable our skills really are outside Claude Code. The other possibility is
fx ask --json as a small step in a script, on a local model through Ollama. We don't have a specific job waiting for it. Install
bash
brew install fx-agent
fx login # or: fx login codex, fx login grok
cd your-project
fx