vercel-labs/fx

 ⠀⠀⠀⠀⠀⠀⣠⣾⣿⣿⣿⠀⠀⠀⠀⠀⠀⠀⠀
 ⠀⠀⠀⠀⠀⢰⣿⡿⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
 ⠀⠀⠀⣠⣶⣿⣿⣷⣶⡶⣶⣶⣆⠀⠀⠀⣴⣶⣶⠆
 ⠀⠀⠀⠉⢹⣿⣿⠉⠉⠀⠘⢿⣿⣧⣀⣾⣿⡿⠃⠀             Tiny, open, embeddable, native coding agent.
 ⠀⠀⠀⠀⣼⣿⡏⠀⠀⠀⠀⠀⠻⣿⣿⣿⠟⠀⠀⠀
 ⠀⠀⠀⢀⣿⣿⠃⠀⠀⠀⠀⢠⣦⠘⢿⣿⣷⡀⠀⠀             curl -fsSL https://fx.sh/setup.sh | bash
 ⠀⠀⠀⣸⣿⡟⠀⠀⠀⠀⣰⣿⣿⠗⠀⠻⣿⣿⣄⠀
 ⠀⠀⠀⣿⣿⠇⠀⠀⠀⠾⠿⠿⠋⠀⠀⠀⠘⠿⠿⠦             ⚠ Status: Experimental. Use at your own risk.
  ⠀⣸⣿⡿⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
 ⣿⣿⣿⠟⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀

fx is a minimal coding agent CLI that still comes with batteries included. It's one native binary written in Zig, so there is no runtime to install, and it runs the same on your laptop, in CI, or inside an agent sandbox. Its output is closer to a Unix shell than an IDE, because the conversation prints into your terminal's scrollback instead of filling the screen.

fx is meant to be embedded and extended. Scripts call it with fx ask, editors connect over the Agent Client Protocol (ACP), and JavaScript apps run the same agent core through libfx with their own tools. Skills, MCP servers, and your project's AGENTS.md add to what the agent knows and can do.

• Small and fast: CI fails a pull request if any benchmarked command, such as fx help or fx status --json, takes longer than 2 ms on Linux (benchmarks). Stable releases for Apple silicon must also fit under a fixed size limit.
• Built-in tools: fx finds, reads, and edits files, runs shell commands, searches the web and reads URLs, and delegates work to subagents. A development server or file watcher can keep running while it works (tools).
• Model choice: sign in to Vercel AI Gateway, use a ChatGPT or Grok subscription, or connect an OpenAI-compatible Chat Completions endpoint such as Ollama or OpenRouter (custom model connections).
• Inline output: you can scroll, search, and copy the conversation like any other command's output, and tmux bindings and copy mode keep working. Terminals that support the Program Status Protocol can also show when fx is busy or waiting for you.
• Your own shell: commands run in your login shell, with your aliases, functions, and PATH. fx runs your startup files once and reuses the result until one of them changes, so a slow .zshrc doesn't slow down every command. Commands also receive AI_AGENT=fx in their environment, so development tools can detect fx.
• Permissions: the default auto mode runs routine work, such as editing project files and running tests, without asking. Anything else goes first to a model that checks the exact action for prompt injection or malicious behavior, and fx holds the action if the check finds a problem. Switch to ask to approve each action yourself, or to full-access to turn the checks off (permissions).
• Sessions: fx saves your conversations under ~/.fx/sessions/, and fx -c picks up where you left off in a project. When a conversation fills the model's context, fx compacts older turns but keeps their full text on disk, where the agent can look it up (sessions).

curl -fsSL https://fx.sh/setup.sh | bash

fx runs on macOS and Linux, on x86_64 and arm64. The installer puts it in ~/.local/bin. fx checks for updates in the background, and fx upgrade installs the latest release right away. CHANGELOG.md lists what changed in each release. You can also try fx in your browser without installing anything.

Sign in with one of these:

fx login          # Vercel AI Gateway
fx login codex    # ChatGPT subscription
fx login grok     # Grok subscription
fx setup          # AI Gateway API key

Then open a project and start fx:

cd your-project
fx

Type what you want done and press Enter. /help lists the slash commands. If something doesn't work, run fx doctor to check your setup.

To send one request without the interactive shell, use fx ask. Add --json when a script needs to read the result:

fx ask "explain the changes in this repository"
fx ask --json "summarize the failing tests" | jq -r .final_output

The libfx README and the libfx docs cover the API. The WebAssembly build is experimental.

The CLI also picks up these extensions:

• Skills: fx loads skills from .fx/skills and ~/.fx/skills, and from the folders other agents use, such as .claude/skills and .agents/skills. The model sees each skill's name and description and reads the full instructions only when it needs them, so installed skills take little room in the context (skills).
• MCP servers: /mcp add slack sets up Slack, and fx mcp manages servers from your terminal. A server's tool schemas load when the model looks for them (MCP).
• Project instructions: fx follows the AGENTS.md in your project, nested AGENTS.md files for the directories they cover, and ~/.fx/AGENTS.md for your own instructions across projects (project instructions).

examples/ has small libfx apps you can run with an AI Gateway API key:

• node-chat: a terminal chat with streamed replies.
• browser-agent: the agent in WebAssembly behind a plain HTML form.
• nextjs-agent: a Next.js App Router form with a streaming route handler.
• nuxt-agent: the same idea in Nuxt, with a Nitro route.

The manual at fx.sh/docs covers models, permissions, configuration, skills, MCP, sessions, and every CLI and slash command.

To learn fx, read the manual as Markdown. fx.sh/llms.txt indexes every page, fx.sh/llms-full.txt puts the whole manual in one file, and adding .md to any docs URL returns that page as Markdown. For the flags of the installed version, run fx <command> --help.

To run fx from your own tools, use fx ask. It takes the prompt as arguments or on stdin, writes its result to stdout and progress and diagnostics to stderr, and exits with a nonzero status when the request fails. With --json, the final answer is in final_output.

To inspect a user's setup, run fx status --json for configuration and runtime information, or fx doctor --json for local health checks. Neither changes the setup.

fx ask uses the user's saved permission mode, which is auto unless they changed it. Pass --full-access only when the user asks for it, because it turns off fx's permission checks.

To change fx itself, read AGENTS.md first, then build and verify the change as described in Contributing.

Run /feedback inside fx to open the feedback form, or open an issue. If something looks wrong in your terminal, /trace creates a private diagnostic report you can attach (share feedback).

You need Zig 0.17.0 to build fx:

git clone https://github.com/vercel-labs/fx.git
cd fx
zig build
./zig-out/bin/fx

Test your changes with the freshly built ./zig-out/bin/fx, because a bare fx may be an older copy on your PATH.

zig build test runs the unit tests, which sit next to the code they cover. The end-to-end tests in tests/e2e/ drive the built binary through tmux and need Bun.

The source is laid out like this:

fx/
├── src/
│   ├── main.zig     entry point that wires everything together
│   ├── core/        agent loop, sessions, config, permissions, MCP, skills
│   ├── builtins/    built-in tool specs, providers, and the system prompt
│   ├── tools/       built-in tools: files, shell, web, skills, subagents
│   ├── ui/          terminal rendering and input
│   ├── gateway/     model provider transport
│   ├── mcp/         MCP v2 engine, behind FX_MCP_ENGINE=v2
│   └── acp/         Agent Client Protocol server
├── sdk/             libfx, the JavaScript library
├── tests/           end-to-end tests and model evals
└── benchmarks/      startup, memory, and latency benchmarks

Read CONTRIBUTING.md before you open a pull request. It covers verification, CI, and PR labels. If a coding agent is helping you, point it at For coding agents.

Report vulnerabilities through the contact page instead of a public issue.

Apache-2.0. Third-party licenses and attributions are in THIRD_PARTY_NOTICES.md.

Interface sounds by cuelume.