Vercel Labs has open-sourced fx, a coding agent written entirely in Zig that ships as a 6.3 MB native binary with a 10-microsecond cold start. While every other coding agent ships as an Electron app or requires a Node.js runtime, fx takes a different approach: it is designed to be embedded in your own application, not just used as a standalone tool.
The project is available on GitHub under an Apache 2.0 license. It is pre-1.0 and moving fast, but the design philosophy is already clear — and fundamentally different from Claude Code, Cursor, or Codex CLI.
Why Zig, and Why It Matters #
Zig reached version 1.0 in late 2025, which makes fx one of the first high-profile production tools built on the stable release. The choice of Zig is not a stunt. It is the direct reason for the performance numbers Vercel is advertising.
There is no runtime to install. A Zig binary runs identically on your laptop, in CI, or inside an agent sandbox. The binary is smaller than equivalent Rust code because Zig has fewer built-in safety abstractions. At runtime, Zig is only 2% slower than C, compared to 8% for Rust and 34% for Go. And because Zig uses explicit allocators — every memory allocation is passed as a parameter — the memory profile is both small and predictable. That 10-microsecond cold start is not a marketing claim: it is an architectural consequence.
This matters for a specific class of use case: spawning agents programmatically inside larger orchestration systems. When you need to start an agent 50 times a second in a pipeline, cold start time stops being trivial.
Three Ways to Use fx #
fx exposes the same agent engine through three integration surfaces:
- CLI —
fx ask "..."runs from the terminal. Unlike most agent CLIs, fx writes its conversation into ordinary scrollback rather than taking over the terminal with a persistent interface. - Script component — Call it from CI pipelines, makefiles, or build steps. No daemon, no background process — invoke and exit.
- Embedded library —
createFxAgent()loadsfx-core.wasmin a JavaScript host.createFxTerminal()embeds the full interactive terminal viafx-term.wasm. For editor integration,fx acpspeaks the Agent Client Protocol over stdin/stdout. Vercel also ships a@ai-sdk/harness-fxadapter for direct Vercel AI SDK integration.
One design decision stands out: an embedded fx agent does not inherit the CLI’s shell, filesystem, or stored credentials. The host application provides the permissions model, session storage, and network transport. You give the agent only the tools it needs. That is the right architecture for embedding AI into a product that real users will touch.
The Extension Model #
Beyond the core surfaces, fx supports MCP servers for connecting external tools, subagents for delegating parallel work, and skills for reusable capability modules. A reviewer model can gate what the embedded agent is permitted to do — a meaningful control surface when you are shipping agent behavior to customers.
Supported AI providers at launch are Vercel AI Gateway, ChatGPT, and Grok subscriptions. Notably absent: Anthropic direct, Azure OpenAI, and Gemini. If your team is committed to Claude models, you would need to route through Vercel AI Gateway, adding a dependency and a cost layer.
Who This Is For #
fx is not competing with Claude Code or Cursor for developer mindshare. Vercel says this plainly: fx is aimed at developers building agent infrastructure, not people looking for another chat window.
The practical use cases are teams building an internal coding interface, a research harness, or a specialized agent inside a larger product. The WASM build means the agent can run in the browser without any server-side component. For teams that have been stitching together LangChain orchestration or OpenAI client wrappers to embed agent behavior, fx offers a native-binary alternative with a much smaller footprint.
What to Watch Before Adopting #
Pre-1.0 status is a real warning. The API is already moving: createFxAgent() was recently renamed to createFxEngine() in current documentation. There is also a name collision with the existing fx CLI tool for JSON manipulation — expect confusion in search results and package managers. And the April 2026 Vercel security incident is worth reading before adding any Vercel Labs project to a production supply chain.
None of this disqualifies fx. But “pre-1.0 and changing quickly” is a statement you should weight appropriately before building a product dependency on it.
The Broader Shift #
The coding agent market has been a race to ship the most capable chat interface. fx is a bet that the next phase is not about which agent is smartest — it is about which agent can be embedded most cleanly into the products developers are already building. A 6 MB binary with a 10-microsecond cold start is an architectural statement. Vercel is positioning fx as infrastructure, not software. Whether the bet pays off depends on whether the API stabilizes before the space consolidates.
The WASM SDK documentation and the InfoQ writeup are the best places to start if you are evaluating fx for a project today.