Home / Articles

Skipper: A Local Dashboard for Running Many Claude Code Sessions at Once

A free open-source dashboard that tracks Claude Code sessions, subagents and loops, pings you on permission prompts and lets you steer agents from a browser or phone.

Publisher: REACTAPP.TOP6 min read1433 words

This article is published in English.

Sources: Claude Code Dashboard: How I Track Multiple Claude Code Sessions, Subagents and Permission Prompts in One Place

The terminal layout changes once Claude Code stops being a one-session habit. Four or five sessions run side by side: one refactoring a backend, one writing tests in a git worktree, a /loop asleep until CI finishes, and another that spawned three background subagents and is waiting on them. It feels productive until a pattern appears. Coming back to a terminal after twenty minutes, the session turns out to have been sitting there the whole time, asking whether to allow a command. That is twenty minutes of an AI coding agent idle because it needed one keypress while attention was on a different window.

Running several agents in parallel is powerful, but terminals were never built to keep track of them. The question that matters is which session needs a person right now. No existing tool answered it, so Skipper was built.

What Skipper is

Skipper is mission control for Claude Code. It is a free, local, open-source dashboard that lists every session, subagent, /loop and todo list on the machine, live, and lets you steer them without hopping between terminals. One command starts it:

npx github:bilol-makhmudov/skipper

Open http://localhost:4317 and each session appears with a clear state:

  • Needs permission: Claude is waiting for approval to use a tool.
  • Needs you: the turn is finished and the next message is up to you.
  • Working: the session is running now, with the tool in use and the turn's elapsed time.
  • Sleeping: a /loop is waiting for its next wakeup, shown with a countdown and the reason it paused.

Sessions that need attention float to the top. The tab title shows how many sessions are waiting, and the favicon switches to amber or red, so a pinned tab still gives the signal.

Watching subagents, todo lists and loops

Past the session states, the dashboard exposes what each agent is doing:

  • Live todo and task lists, plus the task lists of agent teams with their owners and blockers.
  • Background subagents with their name, model, worktree branch and the time of their last action.
  • /loop timers that show the next wakeup and the reason for the pause.
  • Workflows, pull requests and artifacts, one click away.
  • A feed of everything that happened while you were away, across all sessions.
  • Token usage by day, project and model, subagents included, with CSV export.

Notifications for permission prompts

This is the feature Skipper exists for. One command sets it up:

skipper hooks install

It adds two official Claude Code hooks, Notification and Stop, to the settings. After that, the instant a session asks for permission there is a chime and a desktop notification that stays until clicked, showing the exact command waiting for approval. A finished turn gets a softer chime.

The hook does almost nothing. It appends a line to a local event file and exits, so it can't slow a session down or break it, which counted for more than any feature. Anyone who doesn't want a browser tab open can run skipper hooks native on to send system notifications directly.

Steering, not just watching

  • Message a session. For a session open in a terminal, the text goes into that very conversation. A finished session is revived in the background, from its original working directory.
  • Edit tasks. They can be added, renamed, completed or deleted, although a task that still blocks another cannot be removed.
  • Private notes. Each session can hold notes, and one can be forwarded to Claude when the time is right.
  • Read-only mode. The --read-only flag turns off every write.

Claude Code on a phone, with a voice assistant

Once the dashboard worked, the next wish was to use it away from the desk, say while making tea and wondering whether a migration has finished. For that, Skipper has an iPhone app built with Flutter. Pairing happens by scanning a QR code from the dashboard, and the app shows the same sessions, activity and conversations. The voice assistant is the most used part.

One tap on the mic and the question "What needs me?" gets a spoken reply: which session awaits permission, what finished and what is still running. A follow-up works without repeating the session name. Saying "tell the checkout session to stop after this test" makes the assistant read the message back and wait, and nothing is sent until a yes. It handles English and Uzbek. In hands-free mode it keeps listening and stops talking the moment the user starts.

Under the hood the assistant is Claude Code itself, run headless with a small set of read-only Skipper tools exposed over MCP, plus one tool for proposing a message. It sees everything Skipper sees and changes nothing without confirmation.

How Claude Code stores sessions on disk

The pleasant surprise was that Claude Code already writes everything a dashboard needs under ~/.claude:

  • ~/.claude/projects/*/<session>.jsonl holds session transcripts: prompts, replies, todo lists, subagents, loops and pull requests.
  • ~/.claude/sessions/*.json records which sessions are live and whether each is busy or idle.
  • ~/.claude/tasks and ~/.claude/teams hold task lists and agent teams.

Transcripts are read incrementally from the last position already processed, which keeps even very large histories responsive. Changes travel to the browser over Server-Sent Events as file watchers fire, so there is no polling and no refresh button. The backend is plain Node.js without runtime dependencies, and the front end is ordinary HTML, CSS and JavaScript with no build step or framework.

Local-first and private

Sessions contain code and prompts, so a tool that reads them should be dull about security.

  • No outbound traffic. The tool ships without telemetry, a CDN or any external call. Only the optional spoken replies on the phone go online, through Microsoft's voices, whereas the speech-to-text step is handled on the machine by Whisper.
  • Bound to loopback by default. Host header validation guards against DNS rebinding.
  • A token gates phone access. On a local network the startup output includes a link carrying a random token, and requests without it get nothing.
  • Write paths are hardened. Each one demands a custom header, a JSON payload, a same-origin request and IDs that pass validation. The claude CLI receives messages as an argument list and never through a shell.
  • Strict Content Security Policy. All content is rendered as text and never as raw HTML.

Installing it

npm install -g github:bilol-makhmudov/skipper
skipper hooks install      # chime and notify on permission prompts
skipper service install    # run in the background, start at login

Adding --demo --open opens a tour with sample sessions. Node.js 20 or newer is required, and macOS, Linux and Windows are supported.

Common questions

  • Is there one view for every session? Skipper discovers sessions under ~/.claude by itself, and a new session shows up as soon as it starts.
  • Are subagents, agent teams and worktrees covered? They are. Background subagents, team task lists with owners and worktree branches are all listed.
  • Is there a ping when permission is needed? After skipper hooks install, permission prompts, questions and completed turns each trigger a chime and a desktop alert.
  • Can a message reach a terminal that is already running? It lands in that terminal's own conversation. A session that has closed is resumed quietly in the background.
  • What about cost and status? The tool is free and MIT-licensed. It is an independent project and has no connection to Anthropic.

On the roadmap are encrypted phone connections, an Android build and more actions on sessions from a pocket.