Fleet 1.13 is herePipeline orchestration, Brain daemon, and 136 agent templates.Read the release notes →
Fleet

Getting Started

Get Fleet installed and your first agent running in minutes. This guide covers prerequisites, installation, and essential concepts.

Prerequisites

Fleet is a single binary. No runtime dependencies, no Docker, no Node.js. You need two things:

tmuxRequired — agents run as tmux sessions
tmux -V
gitRecommended — for version control and GitHub integration
git --version
Optional: gh CLI for GitHub integration. Run fleet doctor to verify all prerequisites.

Installation

Download the latest binary for your platform:

curl -fsSL https://fleetctl.ai/install | sh

Or download directly from the downloads page.

Verify the installation:

fleet doctor

Quick Start

Follow these six steps to go from zero to a running agent fleet:

1

Initialize your project

Create a .fleet/config.yaml in your repository root.

fleet init
2

Define your agents

Edit .fleet/config.yaml to add agents with roles, departments, and prompts.

# .fleet/config.yaml
product:
  name: my-app
  description: "A web application"

agents:
  - name: frontend-dev
    role: developer
    department: engineering
    reports_to: tech-lead
    prompt: "You are a frontend developer..."
    subscriptions:
      - ticket_ready
3

Check fleet status

View the Fleet status dashboard.

fleet status
4

Start an agent

Select an agent from the sidebar and start it, or use the CLI.

fleet agent start frontend-dev
5

Start the watcher

Enable autonomous operation. The watcher polls GitHub labels, processes subscriptions, and fires scheduled agents.

fleet watcher start --supervised
6

Monitor your fleet

Check fleet health and view unified logs across all agents.

fleet status
fleet log --since 1h

Core Concepts

Understand the six pillars of Fleet:

Agents

Autonomous AI workers that run in tmux sessions. Each agent has a role, department, and prompt. Status is determined at runtime via tmux, not stored in the database.

Pipelines

Multi-stage workflows with approval gates. Stages run sequentially or in parallel with automatic dead session detection and rejection routing.

Fabric

The shared event bus for cross-repo coordination. Agents publish and subscribe to structured events like pr_created, ticket_ready, and pr_approved.

Brain

Event-driven daemon that handles eval scoring, risk assessment, alerts, quarantine, and evolve. Listens on a unix socket with a 5-minute heartbeat.

MCP

Model Context Protocol server exposing 26 tools over JSON-RPC 2.0 via stdio. Enables Claude Code and other MCP clients to manage the fleet programmatically.

Org Hierarchy

Two-tier architecture: org agents (CEO, CPO, CTO) see all repos via org.yaml. Repo agents are scoped to a single project via .fleet/config.yaml.