The real terminal
for the AI Native era.

Agentty is a native terminal app for running AI coding agents such as Claude Code, Codex, Gemini CLI, Amp and OpenCode. It shows what each agent is doing, runs as many agents as you want in split panes, and lets sessions share context with each other.

Claude Code and Codex have the deepest integration, including live status, session management and context sharing. Gemini CLI, Antigravity CLI, Amp, OpenCode, Copilot CLI, Cursor Agent, Qwen Code, Droid, Goose, Crush, Aider and local Ollama models are detected and can be launched from the app.

tmux splits a terminal into panes but doesn't know what runs inside them. Agentty also gives you unlimited splits, and on top of that it knows when an agent is working or waiting for you, manages Claude Code and Codex sessions, passes context between them, and includes an in‑app browser, Git and usage views.

Both are native terminals for AI coding agents with split panes, attention notifications and an in‑app browser, but cmux is macOS‑only while Agentty also runs on Windows and Linux. cmux is built on Ghostty and focuses on vertical tabs and a scriptable CLI and socket API. Agentty goes further into the sessions themselves: it links Claude Code and Codex sessions so context flows between them live or both ways, lets you search, resume and move sessions between the two, and adds an AI process monitor, a usage dashboard, a built‑in Git page and MCP server management.

Yes, Agentty is free to download. The source code is public under the Apache License 2.0.

No. Session lists, usage numbers and handoffs are built from files on your computer.

Yes. Build my idea takes a description, a plan or files, creates a project and starts Claude Code or Codex to build it with a live preview in the in‑app browser. The built‑in Launch plugin then creates a GitHub repository and a Vercel production deploy with a public URL, and can connect Supabase.

On macOS, yes, on the sites it declares and you approve. It can open pages there — out of sight unless you want to watch — and read them, click, type, fill forms and upload files. You sign in once in the page yourself; Agentty then keeps that site's session cookies in the Keychain and keeps them fresh, so an automation does not begin by signing in again, and on macOS 14 or later it can keep several accounts of the same site apart. The plugin itself never receives your passwords or cookies. It types only into terminals it opened — into yours only right after you use it — and never answers a question or an approval for you. An agent it starts on text read from the web can be run files‑only, with no shell, web or browser tools, so instructions hidden in a page have nothing to reach. Windows and Linux have no in‑app browser yet, so these calls are unavailable there.

macOS 13 or later on Apple silicon, Windows 10 version 1809 or later on x64, and Linux on x86_64 as a .deb (Debian 12+, Ubuntu 22.04+) or an .rpm (RHEL 9+, Fedora). Intel Macs are not supported. The in‑app browser, the browser tools agents and plugins drive, remote access, the menu bar item, mini mode and click‑to‑focus notifications are macOS‑only for now; everything else works on all three.

The Windows installer is not code‑signed yet, so SmartScreen may say the publisher is unknown. Choose More info → Run anyway. It installs for your user only and needs no administrator rights. If your browser refuses to download the .exe at all — Chrome blocks unsigned executables — take the .zip instead; it holds the same installer.

On macOS and Windows, yes: it downloads the new version, verifies its checksum, installs it and starts again. On Linux it tells you a new version is out and links to the release page — you install the .deb or .rpm with your package manager.

Run Help → Diagnose problems, which is also in Settings → System check. It looks for what usually causes this: a CLI a new terminal cannot find because PATH was never set up, several copies of one CLI installed at once, a Claude Code or Codex settings file with a syntax error, variables in your shell startup files that confuse Claude Code, and terminal integration files that are missing or out of date. Every finding has its own Fix button, with Fix all beside them, and a file a fix changes is copied next to the original first. When several copies of a CLI are installed, Agentty starts the newest one.

Its fullscreen screen has turned itself off, which happens after earlier launches ended early. Help → Diagnose problems finds that and puts it back.

Remote access puts your workspaces in a browser on another device — a phone, a tablet or another computer — over your own Tailscale network: every workspace, tab and split pane with its status, the terminal live, typing into it, quick buttons to answer an agent, a prompt box, and browser notifications when a session finishes or asks something. It is never on the public internet: only your own tailnet reaches it, only the Tailscale account that owns the Mac is let in, and then a web password you set yourself. It is macOS‑only for now, and it arrived in v0.2.11. Since v0.2.14 you can also speak a prompt instead of typing it: tap the microphone and the clip is transcribed on your own machine by a local model you install once from Settings → Remote access, so no audio is sent anywhere. A chat workspace opens on the page as its chat, too.

Yes — that is what a chat workspace is for. You talk to one lead agent in a chat, and it plans the work and starts workers, each with its own terminal and its own git worktree on a branch of its own: four at a time, reviewers included. A reviewer reads a worker’s changes before the lead merges them, and the chat itself runs on a branch of the project (agentty/chat-…), so your working folder is left alone until you ask for the result — the lead then opens a pull request or merges. The chat’s bar shows how much of your plan the team is using and warns you before the limit stops everyone midway. It works from the remote page as well.