Tokenman中文

Self-host · Quickstart

Goal: Tokenman installed on one computer of your own and your first session open, in about 10 minutes. With a hub, install from the hub; with none, this one computer works on its own. When you later want to connect several machines or create accounts for colleagues, see Install and deploy.

What you need

The install script brings its own tmux; there is nothing to install first.

1. Install

From a hub (your team's, or one you run yourself per Install and deploy): what you get is the version the hub signed, and installing, following new versions and starting Claude / Codex never go to GitHub.

curl -fsSL https://hub.example.com/install | sh -s -- --host

Replace hub.example.com with your hub's address.

Note: no hub, this computer only — install straight from GitHub:

curl -fsSL https://raw.githubusercontent.com/verkyyi/claude-fleet/stable/bin/fleet-install.sh | sh -s -- --no-hub --host

--no-hub: connected to no hub (local mode). This road goes to GitHub to install and to follow new versions; when GitHub cannot be reached, fleet update says so (「GitHub 不通，留在当前版」) and stays on the current version instead of waiting.

Once installed, the command is at ~/.local/bin/fleet. If the terminal says it cannot find fleet, open a new terminal window and try again.

When it is done it opens the client right in this terminal. Where there is no terminal (a pipe, CI, a remote script) it says 「没有终端，跳过打开客户端」 (no terminal, the client is not opened) and exits 0 — type fleet in a terminal later to open it.

Every argument and the environment variables for automation (FLEET_INSTALL_HOST=0|1 is --host, FLEET_INSTALL_HUB=0 is --no-hub):

curl -fsSL https://hub.example.com/install | sh -s -- --help

2. Open the client

Type fleet in the terminal and press Enter.

The first time, the list is empty and says No sessions — type a name at the bottom.

3. Open your first session

  1. Type a name, for example hello, and press Enter.
  2. A new session opens on the right, with Claude Code in it. Tell it something, for example "hi, what can you do?".
  3. When it replies, everything works.

4. Remember one key

⌘P (iTerm2; in other terminals and on a phone F12 or ⌃] p) opens the panel: every session is here — pick a row to switch to it; write a sentence on the top line and it becomes a new task for a new session.

Closing the terminal window does not stop your sessions; type fleet next time and everything is where you left it. To quit the client altogether, use fleet quit (the sessions keep running).

5. Let sessions work in a code repository (optional)

In ⌘P choose Add a repo… and fill in <owner>/<repo>. The repository goes to ~/projects/<repo>; from then on, pick it when you start a session, and the session changes code and opens pull requests in its own copy, never touching your main checkout.

GitHub repository names accept either case for task creation and migration; existing configuration files and working directories keep their names. If creation fails, read the specific reason: frozen means migration is still in progress, input/registration errors require checking the selected repo, HTTP 401/403 requires checking the host's authentication and permissions, and Hub outages require restoring the service before retrying. Casing does not require re-adding the repo or manually unfreezing it.

Who dispatches your tasks is your setting, not a repository's: once you move to the hub, your new tasks in every repository are dispatched by the hub, while other people's sessions in the same repository are untouched and never block you. A newly added repository selects nothing of its own; it follows you.

A new account starts on the hub: it is set when the hub puts you on its list (an invite, an approved request, an admin's add), your machines pick it up the first time they connect, and your first task from ⌘P or the task tool is dispatched by the hub — skip the migration below.

If your account predates that and still runs the old way, to hand your tasks to the hub, first save a private migration inventory on any of your hosts:

~/.claude/fleet/bin/fleet-repo.sh authority --dry-run > /private/path/my-manifest.json

(Adding owner/repo only narrows the inventory's lists to that repository.) Review your hosts, active owners and blockers, then in order:

  1. Freeze: --freeze --manifest /private/path/my-manifest.json stops your new starts;
  2. Drain: let your running sessions write their handoff and end, and confirm on their node that the native process stopped (--settle --apply settles this machine's finished start records in one pass);
  3. Apply: save a fresh --dry-run; once it shows no blockers, run --apply with it;
  4. Rollback: --rollback, also after a freeze and a fresh manifest; accepted hub tasks must finish and release execution first.

Unknown, offline or incapable hosts of yours, or owners in an unknown state, block the change. Your other logins follow automatically. Keep manifests private: they contain machine and task references.

This changes task control only. Ticket settings, GitHub code hosting and PR checks stay in place. A completed task can still have a PR waiting for GitHub.

Next