How much does fx cost?

fx itself is free: Vercel Labs publishes it under the Apache-2.0 license, so there is no fee for the binary. Costs come from inference only. You sign in with fx login through Vercel or supply an AI Gateway API key, then pay whatever your model provider charges. Pointing fx at local inference avoids per-token charges entirely.

Is fx ready for production use?

fx is labelled experimental, and fx.sh warns to use it at your own risk because frequent changes are coming. The first public tag, v0.0.1, landed on August 17, 2026, with v0.0.3 following a day later. That makes fx a fit for experiments, sandboxes and embedded research harnesses today rather than for critical production pipelines.

Which models can fx use?

Vercel Labs describes fx as model- and provider-agnostic, suitable for both local and cloud inference. Authentication runs either through fx login, which opens the Vercel authorization flow and saves the session, or through an AI Gateway API key. Inside a session the /models command switches between the models your account or local setup exposes.

What can fx actually do in a repository?

fx runs an approval-gated agent loop over your workspace: it asks before sensitive actions such as writing or deleting files, and /permissions changes how strict that gate is. Skills add reusable project instructions, MCP servers connect external tools, and subagents take on independent pieces of work. Sessions resume later with fx -r.

Vercel Labs · 2026-08-17 · major

fx — Vercel Labs open-sources a 6 MB coding agent written in Zig

fx is a coding agent from Vercel Labs, written in Zig and shipped as one 6.39 MiB native binary. It cold starts in 10 microseconds, works with local or cloud models, and is now open source under Apache-2.0.

fx project card — a tiny, open, native coding agent from Vercel Labs

A tiny native coding agent from Vercel Labs, built to be embedded anywhere a shell can run.

Key specs

Binary size	6.39 MiB
Cold start	10 microseconds

Quick facts

Maker	Vercel Labs
License	Apache-2.0
Language	Zig
Latest version	v0.0.3 (August 18, 2026)
Status	Experimental
Inference	Local or cloud, provider-agnostic
Extensibility	Skills, MCP servers, subagents

What is it?

fx behaves like a Unix shell rather than an IDE in the terminal — it preserves scroll history, prints little, and draws complex TUI very sparingly. Vercel Labs built fx as an internal tool and released it publicly on August 17, 2026 under Apache-2.0. The whole agent arrives as a single 6.39 MiB native binary with no runtime to install first.

How does it work?

Compiling ahead of time with Zig is what makes the speed claims possible: fx performs no unnecessary work or I/O before it accepts input, which Vercel Labs measures as a 10 microsecond cold start and a single-digit-megabyte memory baseline. Installation is one curl-to-bash script that drops the binary in ~/.local/bin. From there, slash commands cover status, model choice, permissions, images and usage.

Why does it matter?

Most coding agents assume a developer laptop with Node or Python already configured, which rules them out of thin environments. A few-megabyte static binary that starts instantly fits CI containers, agent sandboxes, cheap VMs and scripts that spawn a fresh agent per task. Keeping the harness provider-agnostic also means teams running their own inference are not locked to one vendor's endpoint.

Who is it for?

terminal-first developers and teams embedding agents in sandboxes

Frequently asked questions

How much does fx cost?: fx itself is free: Vercel Labs publishes it under the Apache-2.0 license, so there is no fee for the binary. Costs come from inference only. You sign in with fx login through Vercel or supply an AI Gateway API key, then pay whatever your model provider charges. Pointing fx at local inference avoids per-token charges entirely.
Is fx ready for production use?: fx is labelled experimental, and fx.sh warns to use it at your own risk because frequent changes are coming. The first public tag, v0.0.1, landed on August 17, 2026, with v0.0.3 following a day later. That makes fx a fit for experiments, sandboxes and embedded research harnesses today rather than for critical production pipelines.
Which models can fx use?: Vercel Labs describes fx as model- and provider-agnostic, suitable for both local and cloud inference. Authentication runs either through fx login, which opens the Vercel authorization flow and saves the session, or through an AI Gateway API key. Inside a session the /models command switches between the models your account or local setup exposes.
What can fx actually do in a repository?: fx runs an approval-gated agent loop over your workspace: it asks before sensitive actions such as writing or deleting files, and /permissions changes how strict that gate is. Skills add reusable project instructions, MCP servers connect external tools, and subagents take on independent pieces of work. Sessions resume later with fx -r.

Try it

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