How to run 5 Claude Code sessions in parallel without losing your mind
Claude Code is incredible at one task at a time — but most engineers want to scope out three features while one is refactoring and another is writing tests. Here is how Cockpit lets you run multiple Claude Code Agent SDK sessions across projects at once, without context-switching pain.
Most Claude Code users hit the same wall after a week:
"Once I have 3 projects on the go, my terminal is chaos."
You spawn one claude session in project A. Spin up another in project B. Tab back. Forget which one is which. Re-paste your context twice. Your shell scrollback eats half the conversation. Eventually you give up and serialize — one project at a time — and AI productivity collapses to "single-threaded human".
This is exactly the problem Cockpit was built to fix.
The mental model: one cockpit, many flights
Think of each Claude Code session as a flight. With raw claude CLI you are flying one plane at a time. Cockpit puts every flight on a dashboard with named tabs, status badges, and notifications.
Internally each session is a separate Claude Agent SDK process — fully isolated, with its own working directory, its own conversation, its own token budget. Your laptop is the air traffic controller; the AI is the pilot.
Setting up parallel sessions
Install once:
npm i -g @surething/cockpit
cockpit # starts the cockpit at http://localhost:3457
Open three projects:
cockpit ~/work/api-server
cockpit ~/work/web-app
cockpit ~/work/data-pipeline
Each cockpit <dir> adds a project tab to the same cockpit. Switching between them is one swipe / one keypress — no terminal juggling. (The short cock alias works everywhere too — same command, fewer letters.)
Inside each project you can spawn multiple Agent sessions. Common pattern:
| Tab | Session 1 | Session 2 |
|---|---|---|
| api-server | Refactor auth middleware | Write tests for refactor |
| web-app | Implement settings page | |
| data-pipeline | Investigate the prod-export bug |
Each session runs concurrently. When any of them finishes (or asks a question), you get a desktop notification + a red-dot badge on the project tab.
Why this is more than four terminal tabs
Three reasons it beats raw tmux / iTerm splits:
- Notifications you can trust. Cockpit knows when an agent has actually paused for input vs. when it's still working. A red dot only shows up when you are the bottleneck.
- Cross-project session browser. Cmd+K opens a flat list of every running and recent session across every project. "What was that thing I was debugging yesterday?" → one keystroke away.
- Shared shell + bubbles. Each project gets its own xterm.js terminal, plus optional Browser / PostgreSQL / MySQL / Redis bubbles. The agent can drive any of them. So your "test the new auth flow in Chrome" task doesn't need a separate window.
Cost: yes, you'll burn more tokens
Be honest about this. Running 5 sessions in parallel means up to 5× token spend. Two ways to keep it sane:
- Reserve cheap models for "always-on" sessions (e.g.
/qaclarification mode), reserve Sonnet/Opus for the deep work tab. - Use
/qa(clarify-only) and/fx(diagnose-only) modes generously — they don't write code, so they don't compound.
What "20× productivity" actually means
We don't actually believe in 20× productivity from AI. What we do believe is that AI agents are now I/O-bound on you, the human. Every minute you spend re-pasting context, switching terminals, or re-explaining what file you meant is a minute of agent idle time.
A cockpit is just an interface that respects how much I/O bandwidth a human has. Five quiet agents finishing tasks in the background, three coming back to you with questions in priority order — that's the actual upside.
Try it: npm i -g @surething/cockpit · GitHub · Try Online