Glut

curl -fsSL https://github.com/axiomata-labs/glut/releases/latest/download/install.sh | sh

Glut assumes the model can be steered by anything it reads — a README, a web page, a command's output — and limits what it can do instead of trying to detect it. It is a lean tool-calling loop over any OpenAI-compatible, Anthropic or Responses API — OpenRouter, OpenAI, a ChatGPT plan, Anthropic, Gemini, DeepSeek, Mistral, local models through Ollama or LM Studio, and two dozen more. It reads and edits your project, runs your tests and builds, and shows you every step in a terminal UI made for long sessions: several agents side by side, an arena that pits models against each other, plans that run while you sleep, and your editor at hand.

Status: 0.1.0, early. Glut is used every day by its author and is solid on macOS; Linux is tested in containers. Expect rough edges, and please report them.

Why Glut

• Every command in an OS sandbox. macOS Seatbelt, Linux bubblewrap with Glut's own seccomp filter. Commands write your project and the usual caches, nothing else; .env files cannot be read; git hooks, .cargo/config.toml, .vscode/tasks.json and the like stay untouched at any depth; the network goes through a proxy that asks before a new host. See the security model.
• A gate you can read. Reading is free, writing asks — or not, in the mode you choose. A command that only reads (ls, grep, git status …) runs at once; rm -rf, git push and sudo always ask. "Always allow" means the scope you pick: this exact command, git status *, or git *.
• Any model, honestly priced. Two dozen providers out of the box, keys in the macOS Keychain and bound to their host. The footer shows tokens, cache hits and cost as they happen; budgets stop a run before it overspends.
• Several agents at once. Up to seven panes, agents in the background, agents that message each other — within one window and across Glut windows — without ever lending each other rights.
• An arena for models. The same task on two to four models, blind by default, each in a confined copy of your repository; compare time, cost, tokens and diffs, take the winner. A test bench of fixed tasks scores models over repeated runs.
• Built for long work. It compacts its own context ahead of time, never cuts a tool's output (the whole stays in a store the model can page through), snapshots your files before every change (/steps puts them back, /rewind the conversation with them), and resumes a run that was lost midway.
• Your editor, not instead of it. Neovim, VS Code and JetBrains plugins, and the Agent Client Protocol for Zed and others: diffs open in your editor, your answer there is Glut's answer.
• One program, nothing else. Written in Rust, a single binary under 50 MB with everything above inside — no Node, no Python, no runtime to install.
• Eight languages. English, German, French, Italian, Spanish, Chinese, Japanese and Portuguese — the whole interface and the manual.

A look

A first task	The arena	Neovim
failing tests found and fixed, every step shown	three models blind, their time, cost and diff	an edit as a diff in your editor, your change in it

Every scene is a real run; the website replays ten of them as text.

At a glance

As of 2026-11-05, from each tool's own documentation — the full comparison has nine tools, the sources, and where the others are ahead.

	Glut	Claude Code	Codex CLI	opencode	pi
Models	any vendor, local	Claude only	OpenAI, local	any vendor, local	any vendor, local
OS sandbox for commands	on by default	opt-in	on by default	none	none, a container advised
Network of commands	proxy asks per host	proxy, allow-list	off; optional proxy	none	none
Tool output	never cut	cut, rest in a file	token budget	cut, rest in a file	cut, rest in a file
Models side by side	arena, test bench	—	—	—	—
Licence	MIT / Apache-2.0	proprietary	Apache-2.0	MIT	MIT

Install

# macOS and Linux, a prebuilt program (checksums verified):
curl -fsSL https://github.com/axiomata-labs/glut/releases/latest/download/install.sh | sh

# Homebrew:
brew install axiomata-labs/tap/glut

# From source (a current stable Rust):
cargo install glut-tui

More, including what macOS asks the first time and how to build from a checkout: docs/install.md.

Start

glut providers add openrouter        # or anthropic, openai, ollama … — `glut providers catalog` lists them
cd your-project
glut

Type what you want done. ⇧⇥ switches the mode (plan, ask, build, auto), / lists every command, ⌃K shows every key, /help <command> opens its page. A walk through the first session: docs/getting-started.md.

Documentation

Getting started	the first session, modes, the screen
Manual	every command, one page each — also in Glut with /help
Configuration	~/.glut/ and .glut/: settings, providers, MCP, hooks, skills, agents
Security model	what the gate and the sandbox hold back, and what they do not
Architecture	how a turn runs, the crates, where to start reading
Glut as a library	embedding the loop in your own program; glut run's JSON events
Editor protocol	what an editor plugin and Glut say to each other
Glut and other coding agents	nine others compared, dated and sourced — and where they are ahead

Contributing

Issues and pull requests are welcome — see CONTRIBUTING.md. Found a way past the sandbox or the gate? That is the most useful thing you can tell us. A vulnerability goes to SECURITY.md, not to an issue. The repository's AGENTS.md is what a coding agent (Glut, or another) should read before it touches the code.

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option. Unless you explicitly state otherwise, any contribution you submit for inclusion in Glut shall be dual licensed as above, without any additional terms or conditions.