A coding agent for your terminal. Written in Rust, configured and extended in Lua, and small enough to read.

Status: pre-1.0. Things still move.

• A terminal UI built for long sessions. Verb-first tool rows, grouped reads, live output from running commands, collapsible thinking, inline approvals, search, mouse selection, and a transcript printed when you quit. Modeless by default, with a vim mode if you want one.
• Agents. The model can hand work to child agents that run in parallel, or many at once with one swarm call. Each one shows up as a live card, a pinned row above the prompt, and a full view you can open, steer or stop.
• Configured in Lua. ~/.config/kage/init.lua sets options, keymaps, highlights, autocmds and the chrome around the prompt. Plugins add tools, commands, providers and custom blocks, each behind the capabilities it asks for.
• Many providers, one model picker. Anthropic, OpenAI, Gemini, DeepSeek, Z.ai, OpenRouter and other OpenAI-compatible APIs, plus your own endpoints in config.toml. A bundled model catalog knows each model's context window, inputs and thinking levels, and kage models refresh updates it on demand.
• MCP. Tools, resources (@server:uri mentions), prompts as /server:prompt commands, and OAuth login for remote servers.
• Editors over ACP. kage rpc speaks the Agent Client Protocol, so Zed, Neovim and other ACP clients can drive it, with sessions, approvals and agents intact.
• Sessions you can come back to. Every conversation is saved. Resume, clone or fork it, and compact it when it grows.
• Permissions when you want them. Tools run freely by default. Add rules, switch to ask mode, or approve per call, per session or for good.
• Scriptable. kage -p runs one prompt and prints the answer, or a JSON event stream with --json.

kage builds from source on Linux, macOS and WSL. You need Rust 1.88 or newer and a C compiler. Lua is vendored.

git clone https://github.com/QaidVoid/kage
cd kage
cargo build --release
ln -s "$PWD/target/release/kage" ~/.local/bin/kage

With nix, nix develop gives you the pinned toolchain.

kage init                        # starter config, optional API key
kage                             # interactive TUI
kage -p "summarize this repo"    # one prompt, printed answer
kage resume --last               # pick up where you left off
kage doctor                      # check the setup

Keys come from environment variables such as ANTHROPIC_API_KEY, or from kage auth login. Inside the TUI, / opens the command palette, ctrl+p switches models and ? lists every key.

A taste of init.lua:

kage.opt.theme = "kage-dawn"

kage.keymap.set("i", "<F2>", kage.action.OpenModelPicker, { desc = "switch model" })

kage.api.autocmd_create("tool_call", {
  pattern = "shell",
  callback = function(ev)
    kage.log("info", "shell: " .. tostring(ev.data.input.command))
  end,
})

The guide lives in docs/: install, keybindings, commands, providers, Lua config, agents, MCP, permissions, themes, plugins and editor setup. Preview it with cd docs && bun run dev.

cargo build
cargo test --workspace
cargo fmt --all --check
cargo clippy --all-targets -- -D warnings
cargo xtask check-ascii

kage started from studying pi by Mario Zechner. Its agent loop and plugin-first design shaped kage early on.

MIT. See LICENSE.