What's Agent Doing
See what Claude Code is doing at every step, in plain English, live.

Install, in Claude Code:
/plugin marketplace add tzafrir/whats-agent-doing/plugin install whats-agent-doing@tzafrirThen start a new session, or run /reload-plugins. Free and MIT-licensed.
Why
With today's models, Claude goes into deep work: bisects, one-off scripts, long chains of flags. For minutes at a time you see a spinner and commands scrolling past, and you can't easily tell:
- what it is trying to do right now,
- what that command is for,
- whether it's stuck or just slow.
A box above the prompt names the step Claude is on, in the words Claude wrote for it, and how long it has been on it.
A bug fix, step by step
“The cart total is wrong when a coupon is applied. Fix it.” To track it down, Claude runs commands you'd have to stop and decode. The box says what each one is for:
| Claude runs | The box says |
|---|---|
npx tsx -e "import {cartTotal} from './src/cart'; console.log(cartTotal({…coupon:{amount:10},taxRate:0.08}))" | Reproduce the total with a $10 coupon |
git bisect start HEAD v2.3.0 && git bisect run npx vitest run src/cart -t "fixed coupon" | Find the commit that broke fixed-amount coupons · 1m 48s |
git bisect reset && npx vitest run src/cart | Reset the bisect and rerun the cart tests |
The bisect runs for almost two minutes. The box shows it is bisecting, not stuck.
Every step, and how it went
Click ▸ for the history: each prompt, command, edit and reply, with ✓ done, ✗ failed, ⊘ denied or ■ interrupted, and how long it took.
Background agents get a row each
They keep their row after your turn ends, while they're still working.
It only displays
No telemetry, no processes, nothing kept between sessions. It reads the session's activity and draws the box.
Every hook passes its event on as it came. It never allows, denies or alters anything Claude does.
Runs in Claude Code in a terminal and in the desktop app's Code tab, as a mod built on function hooks.