What fx actually is

fx is a coding-agent harness and CLI from Vercel Labs. The site tagline is “Tiny, open, native coding agent.” It is written in Zig. The license is Apache-2.0. The latest tagged release when this was checked is v0.0.5. GitHub showed about 1,990 stars.

It is not a new model. It is not a terminal IDE. The docs want the CLI closer to a Unix shell: small binary, fast start, then skills, plugins, MCP, and subagents on top.

The site marks it experimental. Use at your own risk. They say they will make frequent changes.

Size and shape

fx.sh lists a 6.39 MiB binary, a 10µs cold start, and a single-digit-megabyte memory baseline for fx itself. The pitch is sandboxes and machines that run many instances, not one heavy TUI.

Supported install targets are macOS and Linux on x86_64 and arm64. There is no Windows package in the install docs.

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

The installer puts the binary in ~/.local/bin. You can download setup.sh, read it, then run that copy. The install docs are explicit: the script does not check a signature or a published checksum. GitHub release assets for v0.0.5 do include .sha256 files if you verify the archive yourself. fx upgrade tracks the selected channel.

Build from source needs Zig 0.16 or newer:

git clone https://github.com/vercel-labs/fx.git
cd fx
zig build -Doptimize=ReleaseSafe

The homepage demo runs the full CLI as WebAssembly in browsers with Wasm JSPI. That is a real fx.wasm build, not a fake terminal. The WebAssembly SDK is still experimental.

Auth and models

From a repo:

cd path/to/project
fx login
fx

fx login opens Vercel OAuth and stores the session. fx setup stores an AI Gateway API key instead. v0.0.5 also added fx login codex and fx login grok for eligible ChatGPT and Grok subscriptions. Those OAuth tokens stay with OpenAI or xAI. They are not sent to the Gateway.

The runtime is model-agnostic. Changelog notes keep zai/glm-5.2 as the default for new Gateway sessions. /models lists the active provider catalog. FX_MODEL overrides for one process. /setup switches providers.

fx itself has no license fee. Inference is whatever backend you chose. On Vercel AI Gateway, tokens are billed at the provider list price with no markup. The free tier includes $5 of monthly credit. Buying credits moves the team to the paid tier, and that monthly free credit stops. The free catalog is a subset of models.

Default permission mode is auto. ask prompts on unresolved sensitive calls. yolo turns fx permission checks off. The docs say use yolo only in a workspace and container you are willing to lose. auto can also fire a separate review request on openai/gpt-5.4, so an unresolved call can cost more than the same call in ask.

What you run

The launch directory is the primary workspace. Listing, globbing, search, and reads inside it do not need approval. Writes, deletes, shell, vision, skill install, and paths outside the workspace do.

Noninteractive:

fx ask "explain what this repository does"
fx ask --json "summarize the current changes"

JSON stdout stays parseable. Add --prompt-permissions if a TTY should still get a Y/N prompt.

MCP servers load only from the trusted profile ~/.fx/mcp.json. A cloned repo cannot add a server. Tools are discovered lazily (mcp_search_tools, then mcp_select_tool) so a large catalog does not fill the prompt.

Skills are SKILL.md directories. fx looks at workspace and home compatibility roots (.claude, .codex, .opencode, .agents, .claw, plus .fx/skills). Discovery does not inject instructions. They enter context when invoked. Managed installs go to ~/.fx/skills/.

Subagents get their own model, permissions, and transcript. One-off subagents retire after a result. Persistent ones stay reusable. ctrl+x opens the manager.

Editors talk to the native agent over ACP (fx acp). JavaScript hosts can embed createFxAgent() or createFxTerminal().

What it is not

It is not a finished replacement for Claude Code or Codex. It is a small, still-moving runtime: native binary, optional Wasm, Gateway or a subscription, Unix-like extensions.

If you want a polished daily driver, this is the wrong week. If you want a ~6 MiB agent you can install, script, or embed, fx is the experiment. Read setup.sh before you pipe it.

Source cutoff: fx.sh, the install/auth/permissions/MCP/skills docs, the v0.0.5 changelog, vercel-labs/fx, and Vercel AI Gateway pricing, as retrieved August 22, 2026.