Fleet a side project atPostHog
All packs

Hub

One always-on devbox manages the others. It keeps your devboxes linked and watches their disks.

Add to your Fleet

Workflows

  • Disk watch

    On with the pack

    The hub cleans a devbox whose disk runs low with fleet prune and each tool's own cache cleaner, and names the largest folders only when that cannot free enough.

    Read before you turn it on

    The sweep deletes only this. Through fleet prune: worktrees idle for 6 hours, or 24 when no T3 thread or agent owns them, whose commits are on the default branch or in a merged pull request, whose T3 thread is closed, or whose agent has been idle for a day. Their commits stay on their branch, but their lockfile and tool output changes are dropped. Also worktree folders that Git no longer tracks and no open T3 thread uses. Itself: the oldest part of a T3 boot log over 100 MB. Through pnpm store prune, uv cache prune and go clean -cache: caches those tools fill again. It never runs Nix garbage collection and never removes Cargo build output, the uv cache folder, old T3 versions, Docker data or Bun's cache, which projects can link into. It deletes no other file. For the rest, you decide.

  • Hub links

    On with the pack

    The hub keeps SSH tunnels and T3 links to every devbox open, so each one can start threads on the others.

  • Dream

    Off until you turn it onNeeds GitHub

    Each night, the hub reads the day's agent sessions from every machine and opens one draft pull request that adds to, condenses, and prunes your shared memory.

    Read before you turn it on

    Every run sends your agent sessions to an analyst model and spends your AI plan's usage, so it starts off. Your session text stays on the hub for 14 days.

  • Own-account balancing

    Off until you turn it on

    The hub holds your own Claude or Codex subscriptions in a local proxy, and Claude Code and Codex on the hub, your devboxes, and boxes go through it, spending the Claude subscription whose weekly quota resets first.

    Read before you turn it on

    On the Claude and both plans, the hub keeps your Claude sign-ins in CLIProxyAPI, and every machine in the workflow sends Claude Code through it. On the Codex and both plans, it keeps your ChatGPT sign-ins there, and every machine sends Codex through it. The balancer ranks only Claude subscriptions by reset date. Read Anthropic's terms first, and OpenAI's terms for your plan too. As read on 2026-11-05: the Anthropic Usage Policy, effective Nov 12, 2026, section "Do Not Abuse Our Platform" (https://www.anthropic.com/legal/aup), lists: "Resell, proxy, or otherwise provide access to Claude through unauthorized means, including services that route requests through consumer subscriptions or misrepresent the product or client being used". The Anthropic Consumer Terms, effective Oct 8, 2025, section 2 (https://www.anthropic.com/legal/consumer-terms), say: "You may not share your Account login information, Anthropic API key, or Account credentials with anyone else. You also may not make your Account available to anyone else." Section 3, item 7 of the same terms lists: "Except when you are accessing our Services via an Anthropic API Key or where we otherwise explicitly permit it, to access the Services through automated or non-human means, whether through a bot, script, or otherwise." Balancing is only for your own subscriptions, never for someone else's.

  • Proxy panel

    Off until you turn it on

    The hub's proxy serves its management panel on loopback, with quota history, drain forecasts, and buttons that spend a subscription's banked weekly or Fable reset.

    Read before you turn it on

    The panel can change the proxy's settings and subscriptions, and spend banked resets, which do not come back. It also turns on CLIProxyAPI's plugins, so anyone with the management key can install and run other plugins from CLIProxyAPI's plugin store on the hub; fleet cli-proxy doctor fails on any plugin besides Fleet's pinned quota-ledger. It listens only on the hub's loopback port 8317, which the SSH tunnels also carry to your devboxes and boxes, and it asks for the management key, which stays on the hub. Every tunnelled client reaches the proxy as 127.0.0.1, so 5 wrong sign-ins from any machine ban 127.0.0.1 from the management API for 30 minutes. That also stops the hub's balancer and fleet cli-proxy doctor until the ban ends. Fleet does not route it through the hub's HTTPS.

What it costs

You need one devbox that stays on, and some setup on it. Every other machine keeps working when the hub is down.

Know before you turn it on

The hub holds sign-ins for your MCP servers and reaches every devbox over SSH. Keep it as private as your laptop.

Settings9 settings that your agent asks about
  • hub.aliasWhat is the SSH name of your hub, the always-on devbox that manages the others, such as my-hub?
  • hub.diskSweepGigabytesBelow how many free gigabytes does the hub clean a devbox's disk?Default: 15
  • hub.diskAlertGigabytesBelow how many free gigabytes after a cleanup does the hub tell you?Default: 10
  • hub.httpsHostWhich private host name does a browser on your own computer use to open the hub's sign-in pages over HTTPS, such as hub.internal.example.com? Devboxes always reach the hub through their SSH tunnels. Leave it empty to skip HTTPS.
  • hub.relayPortWhich port on the hub do the SSH tunnels from your devboxes use?Default: 60001
  • hub.dreamSessionsHow many sessions may one Dream run send to an analyst model? Each one spends model tokens.Default: 150
  • hub.threadFolderWhich folder do the hub's threads start in, such as ~/notes? The hub starts its disk alerts there. Fleet adds it as a T3 Code project when T3 does not know it yet.Default: ~
  • hub.proxyAccountsHow many of your own subscriptions must the hub's proxy hold before fleet cli-proxy doctor passes? They are Claude subscriptions, or Codex ones on the Codex plan.Default: 2
  • hub.proxyHistoryDaysFor how many days does the proxy panel keep the quota history behind its charts and drain forecasts?Default: 35

Add Hub to your Fleet

Paste this prompt into a T3 Code thread on a machine that runs Fleet.

Your install prompt

Add this Fleet pack to my setup with me: Hub.

Fleet already runs on this machine. Ask me each choice with your question tool, and run every command yourself.

1. Run `fleet update`, so this machine has the newest engine and packs. Then run `fleet packs`, and leave out each pack that is already on.
2. This pack needs a hub: Hub. Run `fleet hub status`. When it prints "The hub: not named", I have no hub yet. Tell me, leave that pack out, and offer to add a hub and a node first with step 9 of https://raw.githubusercontent.com/Silthus/fleet-engine/main/SETUP.md. Otherwise the name after "The hub:" is the SSH name of the hub. Use it as the host of each SSH command in a note.
3. Read the setup note of each pack. Ask me what it asks, and collect the flags that it names. Where a note names "the setup command" or "the command in step 9", use the `fleet packs` command of this prompt. Do not run a step that a note puts after that command yet.
   - Hub: ~/.local/share/fleet/engine/packs/hub/SETUP.md
4. Turn the packs on with one command, and add the flags that the notes name:
   fleet packs --add hub
5. Run `fleet update` on the hub over SSH, so the hub has the packs before a note's next step needs them.
6. Do the steps that each note puts after the command.
7. Run `fleet doctor`, and explain each line. Do each `To do:` line with me, and say what a sign-in is for before it opens.
8. My other machines take the change at their next `fleet update`. To give it to one now, run `fleet update` there.
Open in T3 Code

Add more packs in one prompt