Human in the loop

Claude hands you the tasks only you can do. They wait until you act, and your answer goes back to Claude.

Install, in Claude Code:

/plugin marketplace add tzafrir/human-in-the-loop

/plugin install human-in-the-loop@human-in-the-loop

Then start a new session, or run /reload-plugins. Free and MIT-licensed.

Source and documentation on GitHub Privacy policy

Why

Claude often needs something only you can do: a key in .env, an approval on staging, a command on your own machine, a decision. Today it says so in the middle of a reply. You're busy steering the next fix, and the ask scrolls away. At best Claude keeps repeating “I still need you to…”; at worst the ask is lost.

With this mod, Claude assigns the ask as a task instead. It waits in the My tasks pane, counted under the prompt, until you do it, answer it or reject it.

An example

Get our Stripe webhooks library ready to open-source.

Claude works through it, and three times hits something only you can do: the integration tests need your Stripe key, making the repo public is your call, and the package needs a name. Each becomes a task, and Claude keeps working:

My tasks

☐ 3 tasks for you

☐ #1 Add STRIPE_SECRET_KEY to .env 2m

The integration tests run against Stripe's test mode and need it.

Done when: It's in .env. Don't paste it here.

d: Done r: Answer… x: Reject…

☐ #2 Decide whether to make the repo public 1m

☐ #3 Choose the package's name on npm now

Esc: back to the prompt Hide

☐ 3 tasks for you · /my-tasks

#1 Something to do

You add the key in your editor, then press d. The key never passes through Claude.

#2 Your decision

1 makes the repo public, 2 keeps it private.

#3 Your words

Press r, type @acme/stripe-hooks, Enter.

Answer whenever it suits you. Answers you give while Claude works wait for it, and your next answer after it stops takes them along. Here Claude gets them as one message, in your words:

My responses to tasks you gave me (Human in the loop): #1 Add STRIPE_SECRET_KEY to .env: Done. #2 Decide whether to make the repo public: I chose "Make it public". #3 Choose the package's name on npm: @acme/stripe-hooks

Claude picks up where it left off: the integration tests pass, the repo goes public, and the package gets its name. Publishing is left to you.

How it works

Claude keeps working. Unlike a question dialog, a task doesn't stop Claude: it carries on with everything else while the task waits for you, and takes the task back if it finds another way.
Nothing scrolls away. The My tasks pane opens by itself when a task arrives, and the status line under the prompt counts what's waiting. /my-tasks opens the pane any time.
Answer your way: Done, an option, an answer in your words, or a rejection with a reason. Long answers, like a log or a stack trace, go through your own prompt box.
Your answer reaches Claude right away while it's idle. While it works, your answer waits, so Claude isn't interrupted, and goes when Claude finishes its turn, together with any task still open, or now with Send now.
Nothing is lost across a compaction or a new session: open tasks stay with the project, and each session keeps its own, so your answer reaches the Claude that asked.

It never draws in the band above the prompt, so it works beside band mods such as What's Agent Doing.

It stays on your machine

No network, no telemetry, no processes. Open tasks are kept in the plugin's own store under your Claude Code configuration directory, by project.
Your answers become part of the conversation, like anything you type. So tasks ask you to put a secret where it belongs and press Done, never to paste it, and an answer that looks like a secret is held with a warning first.