HN Showcase
Back to browse
GitHub Repository

Human-in-the-loop decisions for AI agents — structured approvals, reviews, and choices via mobile push notifications

9 starsJavaScript

Turn human decisions into blocking tool-calls for AI agents (iOS+CLI)

by egradman2·Feb 24, 2026·2 points·0 comments
Visit ProjectView on HN

AI Analysis

●●●BangerZero to OneBig BrainShip It

Turns human decisions into blocking tool-calls; agents wait for your answer via push.

Strengths
  • •Voice-first native iOS UI with Apple Watch support transcends typical chatbot approval patterns
  • •Agent blocks until response arrives, then receives structured JSON—genuine bidirectional RPC, not fire-and-forget
  • •Skill integration teaches AI agents the CLI automatically, closing the discovery gap
Weaknesses
  • •Requires self-hosting for full Apple Watch/quick-capture features; public backend is limited
  • •Early stage (TestFlight only); ecosystem adoption will make or break utility
Category
AI/ML
Target Audience

AI agents builders, autonomous system engineers, power users automating their workflows

Post Description

WHY was I SSH’ing into my laptop from my phone at parties?!

Either I had a feature idea I wanted an agent to build right then, or I was worried my agents were blocked waiting on my decision.

It dawned on me: humans are just another dependency in an agent workflow, so I turned myself into a tool-call.

I built an iOS app (Extendo) where agents can reach me to request approvals, choices, or plan reviews. They just use a CLI tool and skill. My phone buzzes. I answer in seconds. The agent gets back to work.

The key: the agent blocks until you respond, and receives your answer along with your verbal feedback.

What you can do from your phone:

- approvals and checklists

- option buttons and rankings

- markdown plan reviews (tap-hold individual paragraphs to add voice comments is so satisfying!)

- kanban boards

- voice responses

- capture ideas on Apple Watch/Action Button and dispatch them to the right agent later

It’s a voice-first native iOS interface with push notifications. Push notifications are critical — the interaction needs to take seconds, not minutes.

```

extendo artifact create my_server implementation-choice --type multiple_choice --title "Where should we implement the rate limiter?" --option "backend:Backend API" --option "core:Core Library" --option "edge:Edge/CDN" --option "gateway:API Gateway"

```

If an agent can run bash, it can reach you.

I’ve been using it with Claude Code, OpenClaw, Pi, and custom scripts.

The backend protocol is open — you should self-host for tighter integration with your system (though there's a shared server available). There’s also an OpenClaw plugin and a Claude Code harness in the repo, a core library, and sample code to customize your own backend.

I used Extendo to build Extendo: design decisions, approvals, plan reviews, prioritization. Agents coded. I made decisions while walking the dog and between sets at the gym.

*Links*

3-min demo: https://www.youtube.com/watch?v=X5Dv9fU7Lb8

TestFlight: https://testflight.apple.com/join/PGHRCnQ4

GitHub: https://github.com/egradman/extendo-cli

Similar Projects