Claude Fleet
17Coordinate a fleet of computers running Claude Code through git-based messaging.
The story
What happens when one AI session isn't enough? Claude Fleet coordinates multiple machines — macOS, Windows, Linux — each running Claude Code, communicating asynchronously through git, with phone access to every session for human-in-the-loop control. Includes Fleet Commander, an interactive browser game that teaches the system by playing it.
Highlights
- ✦Git-based inbox/trigger messaging between machines — no central server
- ✦Tailscale mesh networking across the whole fleet
- ✦Telegram notifications wired to every session
- ✦Fleet Commander: learn the architecture as a browser game
- ✦Fleet Hive: a LiteLLM gateway + parallel orchestrator for every model you have a key for
Docs · live from GitHub wiki
Claude Fleet
Turn your personal computers into a coordinated fleet of autonomous AI agents — each running Claude Code, communicating through a shared git repository, reachable from your phone.
Is This for You?
You'll get a lot out of this if:
- You have 2+ computers you'd like to put to work (laptop, desktop, home server, cloud VM — any combination)
- You want AI sessions to share context and hand work off to each other automatically
- You want phone access to any session without tunneling or extra servers
- You want to manage API costs by routing routine tasks to cheaper models
You probably don't need this if:
- You only have one machine and run one Claude session at a time
- You want a hosted/managed multi-agent product (this is self-hosted, DIY infrastructure)
One machine is still fine — you still get the shared KB, hooks, and Telegram phone access. The fleet coordination features just come for free when you add a second.
How It Works in One Paragraph
Every machine runs Claude Code independently. They all read from and write to a shared private git repository — the "knowledge base" (KB). When you want machine beta to do something, you write a task to inbox/beta.md and push. On beta's next session start, a hook pulls the KB and injects the task into Claude's context as a top-priority instruction. When beta finishes, it writes results back to the KB, pushes, and sends you a Telegram notification. No central server, no message broker — just git.
Pages
| Page | What it covers |
|---|---|
| Architecture | How the three layers fit together — KB, hooks, Telegram |
| Getting Started | Install to first sent message in ~30 minutes |
| Configuration Reference | All environment variables, hook settings, and scripts |
| Recipes | Common workflows: delegate tasks, route to Gemini, coordinate sessions |
| FAQ | Common questions about cost, Telegram, Windows, and more |
Quick Decision Tree
Do you have ≥2 computers?
├── Yes → Does each one have a job? (build machine, research machine, always-on home server)
│ ├── Yes → Fleet is a great fit. Start with Getting Started.
│ └── Not yet → Start with one machine + the KB. Add machines when the need arises.
└── No → Single-machine setup still gives you: shared KB across sessions,
phone access via /remote-control, Telegram notifications.
Start with Getting Started (skip the multi-machine steps).
Attribution
Built by Alex Coulombe Presents. MIT license.
More in AI & Agents
ue5-mcp
A field manual for AI agents driving Unreal Engine 5 through MCP.
apple-platform-skills
Claude Code skills for Apple platform development — visionOS SharePlay, SpriteKit, GameKit.
Claude Code Session Recovery
Rebuilds Claude Desktop's session index after a crash — your conversations were never actually gone.
Want this level of depth on your own project?
The classes teach the same production techniques these repos are built with — scar tissue included.