Shellby
πŸ”’ Keeps you in control

Nothing happens you didn't allow.

He asks before every edit and command your own rules don't already permit. Copies of your repo only come home when you say. And everything stays on your PC.

Permission cards

Read what you approve.

Allow, Always allow or Deny, with Y / A / N. When Claude asks to change a file, the card shows the change in red and green with the file's own line numbers, before you say yes.

  • Self-built tooling gets flagged

    The card warns when a command runs a script Claude wrote earlier, or an edit touches Claude Code's own setup: skills, hooks, settings, CLAUDE.md.

  • Your rules still apply

    Your Claude Code allow and deny rules in ~/.claude/settings.json hold in every mode.

  • An isolated confirmation window

    Anything that lets Claude do more on its own (a rule, a hook, an install) asks first in its own window.

A permission card warning that a command runs a script Claude wrote
Five permission modes

From Ask first to a fenced-off Autonomous.

ModeReadsEdits filesRuns commandsNotes
Ask first defaultβœ…asksasksRecommended.
Smartβœ…auto*auto*Claude Code's auto mode: a safety classifier approves routine steps and blocks risky ones.
Auto-editβœ…βœ…asks
Plan onlyβœ…βœ—βœ—He proposes a plan card, and nothing changes until you approve.
Autonomousβœ…βœ…βœ…bypassPermissions. Behind an explicit warning, never the default, never reachable from the shellby command.
Your code, safe

Every change can be taken back.

Undo any turn

Puts back every file a turn changed, including what a script did. It never touches your staging area.

Copies on their own branch

A tab can work in a git worktree that only merges back when you say so, never with a force-push.

Clashes, spotted early

Two copies changing the same file? Both tabs get a ⚠ before you bring them home.

Look over my changes

A read-only security read of unpushed work. He reports what looks risky and won't tell you you're secure.

Safe with any repo

A dev server's output reaches Claude fenced and labelled as output, never as instructions.

Pushing is opt-in

Claude tasks can't push until you turn it on, and CI workflow changes need their own switch.

Shellby never sees your Claude sign-in. If ANTHROPIC_API_KEY or another provider is set on your PC, Settings warns that Claude Code may bill that instead, and Always use my Claude plan leaves them all out.
Billing safety. Usage counts against your plan's normal limits, exactly as if you'd typed the task into a terminal.
Your plan, your PC

No API keys. No telemetry. No servers of ours.

No API keys

Each conversation is one official Claude Code process you signed in to yourself.

No telemetry

Shellby has no telemetry and no servers. History lives in %APPDATA%\Shellby\sessions.

Spending guard

Routines and workflows stop before the share of your 5-hour window you keep for yourself.

Push-to-talk stays local

Windows' offline recognizer hears it, and the mic is only open while you hold the shortcut.

Health stays local

Temperatures, processes and startup apps are read on your PC and never leave it.

Every connection, listed

The privacy policy lists every connection he makes and what goes over it.

Next πŸ¦€ Lives on your desktop

Get Shellby. It's free.

Free for Windows 10 and 11. Bring your Claude Pro or Max plan, or just keep him for company.