Parallel tabs
Each tab is its own Claude Code process. Ctrl+T, Ctrl+W and Ctrl+Tab work like a browser.
The same official Claude Code you signed in to, with helper crabs, tabs that don't collide, undo for any turn and checks he runs himself. Nothing to migrate, nothing to give up.

When Claude sends out a subagent, it walks out of his shell as a helper crab: its own lane in the panel and its own crab on your desktop.
Each subagent gets its own lane and its own crab. Click a helper on the desktop to jump to its conversation.
Every agent type becomes a lasting crew member with a name, a hat and a record: runs, failures, tokens and findings Claude acted on. They level up and earn hats.
Each lane shows its task, live activity, tool count, tokens and time. A helper's permission card shows up in its own lane, labelled with which crab is asking.
Each tab is its own Claude Code process. Ctrl+T, Ctrl+W and Ctrl+Tab work like a browser.
A tab can move into a git worktree on its own branch the first time it changes a file. Bring it home merges it back, never with a force-push.
When two copies change the same file, both tabs get a โ and he says so once, before you bring them home.
Each turn ends with the files it changed. Undo puts them back, including what a script or npm install did, and never touches your staging area.
Every step that changes files gets a checkpoint, so you can roll back to just before any one of them.
Branch any message into a new tab with the files exactly as they were then. Run two approaches side by side and keep the one you like.
/tries 3 fix the flaky login test sends the same task to three copies at once, runs each one's tests and ranks them.
Split (Ctrl+\) puts up to four conversations on screen, or pops one out into its own window.
A tab that finished a turn with changes waits under Ready to review until you've looked.
Run checks on any turn runs the project's own tests (plus a typecheck or build if there is one) and stamps the verdict on that turn. Optionally after every turn, and before a copy comes home.
โ
Checks passed ยท npm run test 12s, or โ 2 failing with their names. Works with npm, cargo test, go test and pytest.
With a dev server up, he photographs the page as the turn starts and again when it ends. Slide between them on the turn's changes.
Critical hit! when one turn takes a red suite to green; Clean landing when a copy comes home green on the first try. Never for skipping or deleting tests.
Correct Claude the same way twice in one project, and a card offers to make it a rule in that project's CLAUDE.md, in your own words.
Click a line number (Shift+click for several) and say what should change. Comments collect across files and turns and go back as one follow-up.
Hold Ctrl+Alt+Space and say the task. Windows' offline recognizer hears it, on your PC.
Stop, undo, rewind, try it another way, bring it home, compact, change mode, start the dev server, run a routine.
Enter queues a message while he's busy, and Claude reads it at his next step, so you can change course mid-task.
All of Claude Code's own commands, each with what it does. Esc Esc, ! and @ work like the terminal.
Back after an hour? A recap of what finished, what failed and what's waiting on you.
Every skill, agent, command, MCP server, mod, hook, rule and CLAUDE.md, with an editor for each. Whatever you set up here works in your terminal too.
Every plugin in your marketplaces, most popular first. Each install asks first in an isolated confirmation window.
Save what you ask for again and again, then type /review or run shellby do @review.
Claude Code mods with a switch each, saying in plain words what they can do.
Save snippets, workflows, hooks and rules as .shellby/team.json; anyone opening the repo can set up the same way.
How many tokens a new conversation carries before your first word: Claude Code's prompt, every plugin's listings, your CLAUDE.md.
What each plugin adds to every conversation. Unused for 21 days? Marked idle, with a Turn off that's easy to undo.
this turn: 18k tokens ยท ~3% of your 5-hour window ยท 41% of context.
"This usually takes about 20% and you've 12% left." Send anyway, or send after the reset.
Past 60% full, he offers Compact or Start fresh with a summary while there's still room.
A dot says whether this conversation's cache is warm, cooling or cold, and what that means for your next message.
With the plugin he reacts to Claude Code in VS Code, Cursor, Windsurf, Zed, JetBrains and Windows Terminal, and sits in its status line. Continue in a terminal carries a conversation on with claude --resume, and /shellby:handoff brings one back.
$ /plugin marketplace add x-salmon/shellby $ /plugin install shellby@shellby # status line, right under the prompt: ๐ฆ๐จ Shellby working ยท Lv 5 Claw Coder โฐโฐโฐโฑโฑ ยท ๐ฅต GPU 84ยฐC ยท +25 XP
src/x.js:42 in a reply opens in VS Code, Cursor or Windsurf at that line.
Any file's diff opens in VS Code's own side-by-side view, editable when the file is unchanged.
Fetches, merges anything the remote has (never a rebase), then pushes. It never force-pushes.
A read-only security read of work you haven't pushed, worst first. It changes nothing.
Work with Claude Code on a server or another PC over ssh. Claude Code doesn't even have to be on this PC.
Mark a conversation done so the finished twenty stop burying the two you haven't.
Free for Windows 10 and 11. Bring your Claude Pro or Max plan, or just keep him for company.