Getting Started

Get up and running with AgentLoop in a few minutes. This guide covers installing the desktop app, signing in, choosing a model, and running your first project.

AgentLoop is currently in an early access beta. Expect rapid iteration.

Requirements

macOS — Windows and Linux builds are coming soon
Git installed and on your PATH (AgentLoop drives git for parallel worktrees)
Podman (optional, only if you want container-sandboxed agent execution)

Installation

Download the app for your platform:

macOS (Apple Silicon) — Download, then drag AgentLoop into Applications and open it
Windows — coming soon
Linux — coming soon

Once installed, launch AgentLoop from your application menu. The first time you open it, the workspace shell appears with sign-in.

First-Time Setup

On first launch, sign in to AgentLoop:

Click Sign in via Web Browser
Your browser opens to the AgentLoop authentication page
Log in with your account credentials
Return to the app — sign-in completes automatically

The desktop app is in closed beta. If your account isn’t on the allowlist yet, you’ll see a Join the waitlist prompt after signing in. Once you’re approved you can come back and finish setup. macOS builds are Developer-ID signed and notarized.

Step 2: Choose Your Model

After signing in, you’ll be prompted to select an LLM provider to power your agents. AgentLoop supports Anthropic, OpenAI, Gemini, and OpenRouter — connect any of them with your own API key, or sign in with a Claude subscription. The default model if using Claude is Opus 4.8, which you can change any time from the account menu or the chat model picker.

Connecting Anthropic (Claude Models)

There are two ways to use Claude models in AgentLoop. Pick whichever matches what you already have:

Use Claude subscription — sign in with your Claude Pro or Max subscription. This is the quickest path if you don’t have an Anthropic API key, and unlocks Claude models without managing one.
Anthropic API key — choose Connect your own API key, pick Anthropic as the provider, and paste your key.

Either option makes Claude models available in the model picker. The default model is Claude Opus 4.8 (shown as anthropic/claude-opus-4-8).

Open Your First Project

Add a Project

From the workspace sidebar, add a folder on disk that contains a git repository. AgentLoop will index it and remember it for next time.

(Optional) Run the Spec Workflow

For a fresh project, AgentLoop can run a guided spec workflow — an interactive questionnaire that helps you describe what you want to build before any tasks are created. You can skip it and just start chatting.

Describe What You Want

Open the chat surface and type your request in plain language:

Break down “user authentication” into tasks and add them to the board

The Product Manager agent breaks this into tasks with sensible dependencies and lands them on your kanban board.

Watch the Board

Open the Tasks view to see the resulting tasks across the lanes (todo, in-progress, review, done, blocked).

Let the Agents Work

Toggle the runner on for your project. The orchestrator picks up ready tasks and assigns them to the right agents. You’ll see them light up in the Agents view as work begins.

What Happens Next

Once the runner is on:

The orchestrator finds tasks with no unmet dependencies
Each task is assigned to the right agent (engineer, QA, etc.)
Engineer agents work in isolated git worktrees so parallel work never collides
QA agents validate completed work before it moves to Done
The release agent can open pull requests for finished work

You stay in control: pause the runner any time, ask questions in chat, or move tasks around on the board manually.

Configuration

AgentLoop stores configuration in TOML files:

Location	Purpose
`./.agentloop/config.toml`	Project-level config (lives next to the repo)
`~/.config/agentloop/`	User-level config, database, and credentials

Most settings can also be edited from the Configuration panel inside the app. See Configuration for the full reference.

Next Steps

Desktop App — Tour the workspace, panels, and chat surface
Kanban Board — Visual task management
Agents — Meet the specialized agents
Orchestrator — How tasks get scheduled

Need help? Join our Discord community for support and discussions.