fx :Tiny, open, native coding agent.
TL;DR Highlight
Vercel Labs released fx, a 6.66MiB single-binary coding agent written in Zig.
Who Should Read
Backend or DevOps developers who want to embed a lightweight coding agent into CI or sandboxes
Core Mechanics
- 6.66MiB Zig binary
- fx is a coding agent harness (the wrapper that lets a model run tools) and CLI written in Zig.
- One binary installs into your terminal, CI, or agent sandbox.
- Shell-like output, not an IDE
- It preserves scroll history and keeps tool activity inside the conversation.
- Open the full transcript when you need more detail.
- Designed to save context
- Tool schemas are concise.
- Skill instructions and MCP tool schemas load only when needed.
- Embeddable and WebAssembly-ready
- libfx lets you embed the core, add tools, and connect skills or MCP servers.
- It runs in the browser too, with your app controlling storage, tools, and network.
- Provider-agnostic by design
- It is meant to work with local models, gateways, direct API access, or subscriptions.
Community Reactions
- Doubts about the binary size
- One commenter argued an agent loop should fit in 200-300KB, so 6MB seems large.
- Another built the repo with zig build on 0.16 and got a 144MB binary, 44MB after stripping.
- Trouble connecting non-Vercel providers
- One user spent 5 minutes and could not find a way to sign in with their own subscription or API key.
- Several others could not find a generic OpenAI-compatible endpoint option.
- Pushback that only Zig is new
- One commenter said the harness is largely the same, just Vercel-flavored.
- They had built a Go harness that is just as portable and deploys to Vercel's sandbox.
- Hands-on reactions and alternatives
- One user ran GLM 5.2 free on a free Vercel account and found it cool though not very polished.
- Alternatives mentioned: a ~40KB C binary, Rust-based Maki, and Nim-based 3code.
How to Apply
- Install fx in a sandbox
- Run the official setup script in an isolated environment and send one request with fx ask.
- It is experimental, so use a test directory rather than a real project.
- Compare startup time and memory
- Run fx and your current agent CLI on the same task and measure startup time and memory.
- Use the numbers if you need to launch many agents in CI or containers.
- Check provider connectivity first
- Confirm in the docs that your model provider or local endpoint can be connected.
- Commenters reported trouble, so test direct API key login first.
Try It
# Run in a Linux/macOS terminal (isolated sandbox recommended)
curl -fsSL https://fx.sh/setup.sh | bash
# Check help after install
fx --help
# Run a single request (fx ask from the article)
fx ask "{{question}}"Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.