Notiformer

Notiformer is a human-in-the-loop API for AI agents. It pauses code before critical actions (spending money, deleting data) and lets you Approve or Deny directly from your phone.

Comment

Inspiration

Autonomous AI agents are powerful, but giving them unfettered execution rights over real-world side effects—like firing API calls, charging credit cards, or executing DELETE SQL queries—is risky.

Existing observability tools like Datadog or PagerDuty are built for infrastructure ops and alert fatigue, not for pausing runtime execution. We wanted to bridge the gap between full autonomy and absolute control: a simple primitive that lets AI agents pause mid-execution and ask a human for approval right from their phone lock screen.

What it does

Notiformer is a zero-noise, human-in-the-loop API built specifically for AI agents and automation pipelines.

With a single line of code (await n.ask()), an agent pauses execution, triggers an instant native push notification to your iOS or Android device (or Slack/Telegram), and waits for a single-tap Approve or Deny.

Key primitives include:

n.ask(): Approval gates for dangerous or high-cost actions. Pauses execution until approved or timed out.

n.select(): Multi-choice prompts when the agent needs a multi-path human decision (e.g., Retry, Skip, or Abort).

n.event(): Non-blocking, fire-and-forget push notifications for background events, error spikes, or task completions.

n.gate(): Remote feature flags toggled directly from the dashboard to instantly pause agent sub-routines without redeploying.

How we built it

We designed Notiformer around low latency and minimal developer friction:

SDKs & Integration: Developed native SDKs for TypeScript/Node.js and Python, alongside a standard REST API with long-polling/webhook fallback for Go, Rust, and serverless environments (Vercel Edge, AWS Lambda).

Notification Engine: Integrated Apple Push Notification service (APNs) and Firebase Cloud Messaging (FCM) to ensure sub-500ms delivery directly to mobile lock screens with interactive action buttons.

Integrations: Multi-channel routing supporting Slack Webhooks, Telegram Bot API, WhatsApp, and Discord.

Infrastructure: Cloud backend deployed in GDPR-compliant European data centers with Redis-backed pub/sub queues for non-blocking asynchronous state resolution.

Challenges we ran into

INSPIRATION

Autonomous AI agents are powerful, but giving them unfettered execution rights over real-world side effects, like firing API calls, charging credit cards, or executing DELETE SQL queries, is risky.

Existing observability tools like Datadog or PagerDuty are built for infrastructure ops and alert fatigue, not for pausing runtime execution. We wanted to bridge the gap between full autonomy and absolute control: a simple primitive that lets AI agents pause mid-execution and ask a human for approval right from their phone lock screen.

WHAT IT DOES

Notiformer is a zero-noise, human-in-the-loop API built specifically for AI agents and automation pipelines.

With a single line of code (await n.ask()), an agent pauses execution, triggers an instant native push notification to your iOS or Android device (or Slack/Telegram), and waits for a single-tap Approve or Deny.

Key primitives include:

n.ask(): Approval gates for dangerous or high-cost actions. Pauses execution until approved or timed out.

n.select(): Multi-choice prompts when the agent needs a multi-path human decision (e.g., Retry, Skip, or Abort).

n.event(): Non-blocking, fire-and-forget push notifications for background events, error spikes, or task completions.

n.gate(): Remote feature flags toggled directly from the dashboard to instantly pause agent sub-routines without redeploying.

HOW WE BUILT IT

We designed Notiformer around low latency and minimal developer friction:

SDKs & Integration: Developed native SDKs for TypeScript/Node.js and Python, alongside a standard REST API with long-polling/webhook fallback for Go, Rust, and serverless environments like Vercel Edge and AWS Lambda.

Notification Engine: Integrated Apple Push Notification service (APNs) and Firebase Cloud Messaging (FCM) to ensure sub-500ms delivery directly to mobile lock screens with interactive action buttons.

Integrations: Multi-channel routing supporting Slack Webhooks, Telegram Bot API, WhatsApp, and Discord.

Infrastructure: Cloud backend deployed in GDPR-compliant European data centers with Redis-backed pub/sub queues for non-blocking asynchronous state resolution.

CHALLENGES WE RAN INTO

State and Timeout Management: Balancing async blocking in serverless vs. long-running server process environments. If an approving user does not respond within a configured timeout window, the agent must safely auto-deny. The system evaluates whether action time is within the allowed timeout; if exceeded, it automatically executes the fallback Deny response.

Lock Screen Interactivity: Ensuring native interactive buttons on iOS and Android could resolve payload promises directly without forcing the user to open an app or log into a web dashboard.

Accomplishments that we're proud of

Reduced developer setup time to under 2 minutes: npm install notiformer plus 3 lines of code.

Built sub-second notification routing from code execution to mobile lock screen.

Solved the runaway LLM loop problem by letting developers drop human gates before financial or destructive side-effects.

What we learned

Building control tools for AI agents requires designing for frictionless control. If approving an action requires logging into a complex SaaS dashboard, developers won't use it. The ideal human-in-the-loop interface is an interactive push notification on a device you already carry in your pocket.

What's next for Notiformer

First-party native integrations with agentic frameworks including LangGraph, CrewAI, AutoGen, and LlamaIndex.

Team-based escalation policies: if user A doesn't respond within 2 minutes, escalate the n.ask() prompt to user B.

Built With

Share this project:

Updates

Submission history