Why Noxy?
The Problem
AI agents are taking real actions on behalf of your users — but when they hit a decision point, most HITL implementations block the entire agent and wait. If the user is offline, in a meeting, or slow to respond, your agent idles indefinitely. Workflows stall. User experience breaks. And most teams solve this with a fragile Slack webhook they built themselves in a weekend.
The Pain
Whether it's a deployment, a financial transaction, or a critical business action, your agent shouldn't be frozen waiting for a human. And your users shouldn't need to be watching a dashboard to stay in the loop. Existing solutions force you to choose between full autonomy and full control — Noxy gives you both.
The Solution
That gap is exactly what Noxy solves.
| Feature | Without Noxy | With Noxy |
|---|---|---|
| Setup |  Days of custom code per agent |  Minutes with SDK |
| Delivery channels | Custom code required | iOS, Android, Telegram, Web |
| Reliability | No queue — lost if user offline | Store-and-forward — never missed |
| Timeout handling | Manual fallback code per agent | Built-in, configurable per request |
| Agent continues after timeout | Custom code required | Native — agent keeps running |
| Security | Channel reads all context | End-to-end encrypted |
| Multi-agent tracking | DIY per agent | Built-in agent tagging + analytics |
| Maintenance | You own the infrastructure | Managed SaaS |
Noxy + Your Agent Framework
Noxy doesn't compete with your agent framework — it completes it. LangGraph and LangChain already model suspension with interrupt(), but they leave the hard parts to you: how does the user actually see the request, on which device, with what encryption, and how does the agent resume when the answer arrives?
Noxy ships official connectors that plug straight into the frameworks you already use, so you keep your graphs, tools, and orchestration logic — and add multichannel, encrypted, store-and-forward HITL with two lines of code:
LangGraph Connector
langgraph-noxy — drop-in HITL node that pauses the graph and resumes by polling the relay.
LangChain Connector
langchain-noxy — HITL middleware for create_agent tool calls.
Not on LangGraph or LangChain? Use any Agent SDK directly — the connectors are sugar on top of the same Python, Node.js, Go, and Rust libraries.
| Feature | Framework HITL alone | Framework + Noxy connector |
|---|---|---|
| Pause & resume the agent | Yes — native interrupt() | Same primitive, wired to webhook resume |
| User notification | You build push, email, Slack, etc. | Push to iOS, Android, web, Telegram |
| Mobile support | Build your own iOS / Android client | Drop-in client SDKs, decide from anywhere |
| Offline / store-and-forward | Lost if user offline at request time | Queued and delivered when device reconnects |
| Timeout / TTL | Suspends indefinitely until you intervene | Configurable per decision, with auto-resume |
| End-to-end encryption | Plaintext over your channel | Post-quantum, ciphertext-only relay |
| Identity routing | One channel, one device per user | Email, phone, user id, or wallet → all devices |
| Multi-language reach | Tied to the framework's language | Connectors today, plus Node, Python, Go, Rust SDKs |
Built for AI Agents
Official connectors for LangGraph and LangChain drop straight into the frameworks you already use. Building with AutoGen, CrewAI, or your own orchestrator? Use the Agent SDKs directly — same encrypted-decision API, no framework lock-in.
Future-Proof Security
Every decision request contains sensitive context about real actions your agent is about to take. Noxy encrypts everything end-to-end with post-quantum cryptography — protecting your users' data today and against the threats of tomorrow.
True Privacy
No decision context is stored in plaintext on our servers. What your agent sends, only your user reads. No data mining, no third-party access, no exposure — your users' decisions stay between your agent and them.
Unmatched Reliability
Store-and-forward architecture ensures every decision request reaches its destination — even when the user is offline, in a tunnel, or on a slow connection. When their device reconnects, the request arrives instantly. No lost decisions, no frozen agents.