Fleet 1.13:Teams are now shipping 5x more PRs with autonomous pipelines.See what's new →
FleetFleet
Comparison

Fleet vs Crystal: Persistent Agent Teams vs Session-Based Parallelism

Crystal is a free, MIT-licensed local desktop app that runs parallel AI agent sessions in git worktrees. Fleet adds a persistent, event-driven coordination layer with role-based governance, GitHub automation, and audit trails — better suited for teams that want agents running continuously rather than on demand.

Crystal is a free, MIT-licensed desktop app that lets you spin up multiple Claude Code and Codex sessions working on different tasks simultaneously, each in its own git worktree. It runs locally and is great for batch work and parallel exploration, but sessions are independently dispatched and reviewed by hand. (See Crystal's repository for its current status.)

Fleet is designed for continuous, autonomous operation. Agents persist, have roles, communicate through an event bus, and react to external events like GitHub label changes. The system is less about spinning up sessions interactively and more about maintaining a running team that handles your development workflow without a person at the keyboard.

Choose Fleet if

Teams that want a continuously running agent team that manages the full development lifecycle — coding, review, and release — driven by GitHub events.

Choose Crystal if

Developers who want a free, local desktop app to run multiple parallel AI coding sessions for batch or exploratory work with minimal orchestration overhead.

Fleet vs. Crystal: side by side

FeatureFleetCrystal
InterfaceHeadless CLI + daemonLocal desktop GUI app
Role system120+ role templates: dev, reviewer, PM, release-manager, etc.Undifferentiated parallel sessions
Event-driven triggersWatcher daemon reacts to GitHub label changes automaticallySessions started manually in the app
Audit trailFull decision and conversation log per agentPer-session history in the app
GovernanceRun-time budgets, 6-dim evaluation, auto-quarantine risk model, approval gatesNo governance primitives; human reviews each session
Cost / licenseFree tier + $49/slot/mo TeamFree, open source (MIT)

Where Fleet is the better fit

  • Headless and autonomous — the workflow continues unattended without a person reviewing each session
  • Role-based system prevents the reviewer agent from also being the developer, maintaining separation of concerns
  • GitHub label watcher triggers the entire dev-review-release chain without any manual dispatch
  • 6-dimension evaluation plus a separate risk model that drives auto-quarantine protect production branches from high-risk agent actions

Where Crystal is the better fit

  • Free and open source (MIT) — runs entirely on your machine, nothing to pay for
  • Simpler mental model: start a session, review the result, merge — no daemon to operate
  • Polished local GUI for watching and comparing parallel approaches and picking the best result

Pricing

Crystal is free and open source (MIT), running locally on your machine. Fleet's pricing is $49 per agent slot per month (Team tier) with a free single-slot tier.

Do they compete, or coexist?

Crystal-style local session parallelism and Fleet's autonomous team model serve different operational patterns. Use Crystal when you want to explore multiple approaches interactively on your machine. Use Fleet when you want a consistent, ongoing team of agents managing your GitHub workflow autonomously day to day.

Frequently asked questions

What makes Fleet different from tools that just run agents in parallel?

The key differences are persistence, role specialization, and event-driven coordination. Fleet agents have defined jobs, communicate through the Fabric event bus, react to GitHub events automatically, and operate under governance constraints. Local parallel-session apps like Crystal focus on interactive concurrency; Fleet focuses on autonomous coordination.

Does Fleet re-run work when an agent fails?

Fleet does not respawn a crashed agent session. Instead, its event-driven reactive chain re-dispatches the work — when a PR needs another pass, the watcher and PR reconciler route the triggering event to the responsible role again. The recovery comes from the chain, not from the watcher relaunching a dead session.

Run your first agent fleet

One binary. Five minutes. See every agent, coordinate every handoff, and keep a full audit trail of what your fleet did.