Running multiple Claude Code sessions at once is the real unlock of agentic work — and the real problem. The machines were never the bottleneck. Your attention is. Here’s how I run eight at a time: four habits that work in any terminal, and a free one-screen dashboard called claude-dash that watches the whole fleet for me. The script is yours at the end.
On Sunday morning my desk looks calm. One laptop, one cup of coffee, the house doing its usual Sunday noise behind me. The truth is under the hood: eight terminal windows, each one holding a live conversation with Claude that hasn’t ended yet. One is reviewing a client deliverable. One is building a newsletter. One has been quietly reorganizing an archive for an hour. At the moment I took the screenshot below, three of them were working at that exact second — and five were sitting there, patient as dogs at the door, waiting for me to come back.
My newsletter is called The Human in the Loop, and people sometimes ask what that means practically. This is what it means. Not one person talking to one chatbot — one person conducting several of them at once, and discovering that keeping up with the machines is a job about attention, not typing.
The failure mode: a hand raised in a window you’re not watching
Here’s how it fails. You open terminal four to check on a build, and you don’t notice that terminal two went silent twenty minutes ago — not because it finished, but because it stopped to ask may I run this command? and nobody answered. The most capable software I’ve ever used, sitting politely with its hand raised while I look at the wrong window. Multiply that by eight and you haven’t hired a team; you’ve built a hallway of closed doors you have to keep knocking on.
Four habits that work in any terminal
Before any tooling, these four fixed most of it. No downloads required:
One session, one job scope
Not one session per project — one per task. “Draft the contract” and “fix the invoice script” get separate terminals even if they’re the same client. Sessions that do one thing can be checked in one glance.
Name the tab after the job findability
Future-you is going to see twelve tabs at 4pm. contract-draft beats bash — 80×24.
Treat the permission prompt as the real work priority
When an AI asks permission, that’s the moment it actually needs a human. Everything else can wait; this can’t. Your whole job in the loop is to be findable at that moment.
Watch context like a fuel gauge cost & quality
Every session carries a running memory, and it only grows. A conversation that’s been alive for two days is slower, costlier, and vaguer than a fresh one. When the gauge runs hot, land the plane and start clean.
claude-dash: a one-screen control room for the whole fleet
Those habits got me from chaos to mere disorder. Then I did the most natural thing you can do in 2026 when the tool you want doesn’t exist: I described it to Claude, and we built it together. First working version in a day, back in July. It’s been growing ever since, the way a garden does.
It’s called claude-dash — about 1,300 lines of Bash, which is a ridiculous and beautiful thing for a dashboard to be. One screen, four zones. A header that counts the fleet. A table: every session, its project, a little colored fuel gauge of its memory, and a status — working, idle, or the coral one that matters, needs you. Below that, a peek pane showing what the selected session is doing right now — and when it’s asking permission, the question appears there, with the answer one keystroke away. Approve with y. Deny with n. Send an instruction into any session without ever opening its window.
My favorite part is the key marked v. Press it and the dashboard speaks — an actual voice, out loud: “One session needs you. Three working. Total context 1.8 million tokens.” Which means I don’t have to sit at the desk at all. I can be at the kitchen table with the kids and the laptop across the room will simply tell me when something needs a human. I built a baby monitor for the machines. I’m only half joking about how much that image delights me.
How to install claude-dash
The script is free: github.com/jlapin85/claude-dash. You’ll need macOS, Claude Code, and jq (brew install jq). Three lines to a running dashboard:
The whole install — clone, make executable, run. The README covers the keys, the tuning knobs, and the voice.
The first time you jump to a session, macOS will ask once to enable Terminal under System Settings → Privacy & Security → Accessibility — that’s how the dashboard drives your Terminal tabs.
Fair warning, honestly given: it’s built for my desk — macOS, Terminal.app, and it assumes Claude Code. It reads the session metadata Claude Code writes to ~/.claude/sessions/ and the transcripts in ~/.claude/projects/, which Anthropic could reshape any day. It is not a product. It’s a Bash script with opinions. Read it before you run it — it’s one file — then bend it to your setup and tell me what you turned it into.
Why a dashboard, and not just discipline
Every tool on my workbench exists for the same reason: to spend less of me on watching and more of me on deciding. The machines got fast; attention became the scarce thing in the house — and not just for work. The dashboard’s real output isn’t the table or the voice. It’s that I left the desk.
The loop needs a human. It doesn’t need him hovering.
— Joe
