Zag

zag

A composable agent development environment. Built in Zig.

Zag is an AI coding agent where the **window system is the platform**. Splits, focus, and buffers are primitives. Everything above that — from the session tree to how agent responses render to which system prompt a model gets — is a plugin.

Think **Neovim's architecture, applied to AI agents**.

---

Quickstart

# Build (Zig 0.16.0)
zig build

# Run — first boot launches an onboarding wizard
zig build run

# Run tests
zig build test

That's it. No Docker, no Python venvs, no npm install.

---

What makes it different

Neovim-style modal TUI	Binary-tree pane splits, vim h/j/k/l focus, per-pane scrolling, dirty-rectangle ANSI diff renderer. Floating popups too.
Per-pane streaming agents	Every pane runs its own agent thread. Different panes, different models, same session. Cancel a turn with Ctrl+C without freezing the UI.
8 providers, 3 wire formats	Anthropic Messages, OpenAI Chat Completions, OpenAI Responses API (ChatGPT/Codex). Anthropic, OpenAI, OpenRouter, Groq, Moonshot, Ollama — plus OAuth sign-in for Claude Max/Pro and ChatGPT.
Live /model switching	Swap provider or model mid-session without losing context. Persists the pick to config.lua.
Lua plugins + async runtime	Custom tools, hooks, keymaps, slash commands, providers, and prompt packs. Blocking I/O (HTTP, subprocess, filesystem, LLM) runs on a worker pool and resumes your coroutine on the main thread.
Mid-turn steering	Type while the agent is running — your message is queued as a system-reminder interrupt instead of dangling at the tail.
Automatic context compaction	Predictive cascade (Lua structure