The mainstream coding agents — Claude Code, Cursor, GitHub Copilot in its full form — all share a similar profile: they assume a developer machine, a reasonably capable shell or IDE, and enough runtime overhead that startup time is not a meaningful variable. They’re designed for interactive sessions where a developer is sitting at a terminal and a few seconds of initialization is invisible.
Vercel Labs open-sourced fx this week under Apache-2.0, targeting the other end of that spectrum. The binary is 6.3 MB. Cold starts take 10 microseconds. Memory usage at baseline stays in single-digit megabytes. It’s written in Zig and compiles to WebAssembly for broad compatibility.
The design target, in the project’s own words: “closer to a Unix shell than a heavy ‘IDE in the terminal’ TUI.” That means minimal system prompts, no product telemetry, sessions stored locally. With local inference enabled and auto-updates turned off, the tool makes zero outbound connections during a coding session — fully hermetic, in a way that most agent tools are not.
Why Zig
Zig produces small, fast binaries with no garbage collector and deterministic memory behavior. It has no implicit allocations, which means the runtime profile is predictable — useful if you’re embedding fx inside a larger system that needs tight resource accounting. The 10µs cold start is only achievable because there’s no interpreter startup, no module graph resolution, no JIT warmup. The process forks, reads the binary, and is ready.
This contrasts with Python-based agent tooling, where import time alone — loading langchain, httpx, and their transitive dependencies — typically runs into seconds. For interactive sessions that difference is irrelevant. For a CI pipeline spawning an agent per commit, or a larger orchestration system spawning subagents dynamically, it matters.
Vercel’s angle
Vercel runs millions of deployments. They have direct motivation to think about agent tooling that can fit inside serverless functions, edge workers, or CI runners — environments where binary size and cold-start latency are real constraints, not abstract concerns. fx originated as an internal tool at Vercel Labs before this week’s open-source release.
Being model-agnostic from the start (the tool routes to local inference, cloud APIs, or gateway endpoints with equal ease) separates the harness from any provider relationship. That’s a meaningful design choice for a company whose infrastructure sits between developers and their deploy targets — it lets Vercel remain neutral on which AI backend teams use.
The extension model follows Unix conventions: skills, plugins, and MCP rather than a monolithic feature set. The core binary stays small; capabilities are composed in.
What this is and isn’t
fx is described as experimental and aimed explicitly at “research and embeddability.” It is not a complete alternative to Claude Code or Cursor for a developer who wants rich multi-file context, integrated diffs, and inline rendering. It’s a harness — the part that handles context management, tool dispatch, and conversation state — that you’d want to embed inside a system that itself handles the interaction layer.
That niche is real and underserved. Most agent orchestration frameworks are Python libraries that happen to include a CLI. A 6.3 MB Zig binary with predictable resource usage is a credible answer to the embeddability problem that those frameworks don’t address well. Whether the community finds enough use cases to sustain development outside Vercel’s internal needs is the open question.
The source is on GitHub. Worth watching if you’re building systems that compose agents rather than using them interactively.