Install Fleet and run your first workflow

Fleet separates authoring from execution: build the workflow in the dashboard or through a compatible MCP client, then let the locally installed binary run it in your repository.

Prerequisites

You need tmux for agent sessions, git for repository work, and Claude Code or OpenCode authenticated on the machine that runs Fleet. The GitHub CLI is required for coding workflows that create and review pull requests.

tmux -V
git --version
gh auth status
claude --version    # if you use Claude Code
opencode --version  # if you use OpenCode

After installation, run fleet doctor for the complete local check.

First run

These steps use the visual builder. If you prefer, follow the same setup and then build the workflow by talking to your AI assistant.

Install the Fleet binary

Install the official binary. The installer selects the correct build for your platform.

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

Run one-command onboarding

From the repository Fleet will work in, check prerequisites, scaffold local config, and start Fleet.

cd my-repo
fleet up

Open the Workflow Builder

Sign in to the dashboard, open Workflows, and choose New workflow. You can also ask any MCP-connected AI assistant to create the same saved definition.

https://app.fleetctl.ai/workflows

Choose a starting point

Use Weekly brief for document work or Ship a feature for the complete coding flow. The template creates the steps and dependency edges for you.

Bind the repo and trigger

Choose the repository the workflow works in. Start manually, on a cron schedule, or when a backlog ticket receives a label.

Save and run

Press Run for a manual workflow, then open History to watch each step, artifact, approval, retry, and terminal result.

What Fleet created locally

.fleet/config.yaml   # repo, agents, types, corpus, connectors
~/.fleet/fleet.db    # local operational state
~/.fleet/logs/       # worker and agent logs

Workflow graphs are saved in Fleet whether you author them in the dashboard or through an MCP client, then pulled with each run. You do not need to recreate the graph as a pipelines: or subscription block in .fleet/config.yaml.

Core concepts

Workflow definition

A stable saved graph of typed steps, dependency edges, triggers, outcome routes, and governance settings.

Local binary

The Fleet process installed on your machine. It pulls queued workflows and runs their agent steps against your repository through Claude Code or OpenCode.

Typed step

One unit of work with a specific contract: create an artifact, open a PR, review an output, wait for approval, publish, merge, or notify.

Run

One execution of a saved definition. Runs keep status, step history, artifacts, decisions, retries, and provenance.

Outcome route

A configured edge for outcomes such as review flag, approval rejection, or execution error, with a bounded retry policy.

Approval gate

A suspension point where a person reviews the current evidence and approves or rejects before the run continues.

How the coding template runs

Ticket label or Run button
        ↓
Refine ticket
        ↓
Develop → coding agent opens a PR
        ↓
Review ── flag ──→ Develop (bounded fix loop)
        ↓ pass
Human approval
        ↓
Merge → Announce

The code step parks the run while the agent works. The PR resumes it. The approval step parks again until a person decides. No subscription processor sits between those steps. The saved workflow owns the sequence.

Build your first workflow →

Learn the builder, templates, repo binding, and run flow.

Configure Fleet →

Understand repositories, agents, types, and house rules.