Vercel fx Is Open Source: A 6MB Zig Coding Agent Built for Embedding

Vercel Labs open-sourced fx, a coding agent written in Zig that ships as a 6.39MB native binary with a 10-microsecond cold start, targeting embedding in CI pipelines and agent infrastructure rather than interactive terminal use. fx exposes three integration surfaces — the `fx ask --json` CLI, an Agent Client Protocol server via `fx acp`, and libfx, the agent core compiled as a Node.js native addon and an experimental WebAssembly build — and connects to any OpenAI-compatible endpoint with sessions and usage kept local and no telemetry. The release is explicitly experimental at version 0.0.13 under Apache-2.0, and early users flagged that setup defaulted to Vercel AI Gateway despite the model-agnostic claim, while the name collides with an existing popular terminal JSON processor also called fx.

Vercel Labs just open-sourced fx https://github.com/vercel-labs/fx , a coding agent written in Zig that ships as a 6.39MB native binary, cold starts in 10 microseconds, and is designed from the ground up for embedding in CI pipelines and agent infrastructure — not for developers who want another chat window in their terminal. Not Another Cursor That framing matters. The coding agent market in 2026 is crowded with tools competing for your interactive workflow: Cursor, Claude Code, Codex, Windsurf. fx is not in that race. Vercel is explicit about it https://x.com/vercel dev/status/2089828083415355806 : fx targets “developers building agent infrastructure more than people looking for another chat window.” If you want to embed a capable coding agent into a CI job, a custom agent loop, or an automated sandbox, there was no clean, lightweight primitive to reach for. fx is that primitive. The numbers back up the pitch. The binary is 6.39MiB. Memory baseline is single-digit megabytes. Cold start is 10 microseconds. There is no runtime to install. It drops into a container or pipeline the same way a static binary does, without dragging along a Node.js runtime or a Python virtualenv. Three Ways to Integrate fx exposes three integration surfaces. The first is the CLI: fx ask --json returns structured output that a shell script or CI job can parse directly without screen-scraping. The second is the Agent Client Protocol server: fx acp runs fx as an ACP server over stdin/stdout, letting editors and other tooling connect to it as a backend. The third is libfx , the agent core compiled as a Node.js native addon and as WebAssembly. In JavaScript, createFxAgent gives you the full agent loop — model calls, tool dispatch, retries — with your own tools and MCP clients wired in. The Node SDK documentation https://fx.sh/docs/lib/node covers the full API. The WASM build is experimental but it means browser-based agent tooling is on the table. The Zig Choice Is Not Accidental Writing a coding agent in Zig in 2026 is a deliberate signal. The language has no garbage collector, no runtime overhead, and compiles to native binaries with predictable latency characteristics. It also has idiomatic WebAssembly compilation baked in, which is why getting a WASM build of libfx was straightforward. Bun — the JavaScript runtime that became one of the fastest Node alternatives — also chose Zig for the same reasons. The choice says: this is an infrastructure piece built to last, not a Python prototype that got popular before it got rewritten. For context, Zig now builds 2.3x faster than Rust and sits at 43K GitHub stars. It is no longer exotic — it is the language teams reach for when binary size and startup time are non-negotiable constraints. Model-Agnostic, No Telemetry fx connects to any OpenAI-compatible endpoint: Vercel AI Gateway, Ollama for local inference, Grok, OpenRouter, ChatGPT. Sessions and usage stay local. No source code or prompts are shared beyond your chosen inference endpoint. That no-telemetry stance will matter more as enterprise AI governance tightens — but it also makes deployment straightforward in security-conscious environments that would otherwise require a policy exception for every new tool. The Caveats Worth Knowing A few honest friction points. First, there is already a popular fx https://github.com/antonmedv/fx in the wild — a terminal JSON processor with substantial developer adoption. The name collision is real and the community has flagged it. Second, early users reported that setup defaulted to Vercel AI Gateway even though the model-agnostic claim implied otherwise; the default configuration could be clearer. Third, fx is explicitly experimental at version 0.0.13. The InfoQ technical breakdown https://www.infoq.com/news/2026/10/vercel-zig-fx-agent/ documents its current capabilities and limitations honestly. Availability fx is available now on GitHub under Apache-2.0 https://github.com/vercel-labs/fx . Version 0.0.13 ships Ultrafast mode for supported OpenAI models with faster launches and shell calls. If you are building agent infrastructure and a full framework feels like overkill, fx is worth evaluating. Go in knowing what it is and what it is not.