fx
A tiny, open-source AI coding agent that starts almost instantly and fits inside a CI pipeline, script, or sandbox — built for machines to spawn on demand rather than for a human to sit and chat with.
🔗 Visit fx
Description
Most AI coding assistants are built like desktop apps: you open them, wait a moment, and start a conversation. fx flips that around — it's designed to be launched a thousand times a day by other programs, doing one small coding task and exiting, the way a shell command does rather than the way a chat app does.
fx is an open-source (Apache-2.0) coding agent CLI from Vercel Labs, written in Zig for a tiny binary (6-8 MB) and a cold start measured in microseconds rather than seconds. It's model-agnostic, working with OpenAI, Vercel's AI Gateway, xAI, or local models, and extensible through skills, MCP servers, and subagents. The intended use case is embedding it inside CI pipelines, scripts, agent sandboxes, or cheap VMs where spinning up a fresh agent instance per task needs to be cheap and fast — not sitting in an editor sidebar answering questions. As a very young project (v0.0.6), expect breaking changes and thin documentation compared to mainstream tools.
💬 Our review
The short version: fx isn't competing with Cursor or Claude Code for your daily coding chat — it's solving a narrower, more infrastructure-flavored problem (cheap, instant, scriptable agents), and it does that job well for what it is.
Against GitHub Copilot or Cursor, which are built around an interactive human-in-the-loop experience, fx's value is in automation contexts: a CI job that needs an agent to fix a lint error and exit, or a sandbox that spawns dozens of short-lived agent instances per hour, where a multi-second startup and heavier footprint would add up fast. Claude Code covers similar automation ground but is heavier and less purpose-built for embedding. The honest catch is maturity — v0.0.6 means the API and behavior can and will shift, and there's no large community yet to lean on when something breaks. Worth adopting if you're building agent infrastructure and want minimal overhead; not a replacement for your daily coding assistant.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Gratuit et open source (Apache-2.0). Coût réel = le fournisseur d'inférence LLM connecté (OpenAI, local, etc.).
Pros
Binaire minuscule et démarrage quasi instantané
Agnostique au modèle (OpenAI, xAI, local, Vercel AI Gateway)
Open source, soutenu par Vercel Labs, extensible
Cons
Stade très précoce (v0.0.6), changements cassants à prévoir
Documentation et communauté encore limitées
Pensé pour l'automatisation, pas pour un usage conversationnel quotidien