The architecture

The internet runs on HTTP. Imperal runs on ICNLI.

An open protocol (icnli.org, CC BY-SA 4.0) that specifies how AI must understand an operational reality before acting on it. Every action Webbee takes is decided through it, not a framework choice, the actual mechanism underneath.

The five layers

One sentence, five layers, your systems.

What actually happens between you typing a command and Webbee executing it against your real infrastructure.

You

Plain language, from any surface, chat, voice, a text message. You say what you want; nothing about the mechanism underneath changes how you talk to it.

Panel

Your visual and command surfaces across Web Panel, Telegram, Terminal CLI (Webbee Code), and external IDEs via Imperal MCP. Conversations, cards, dashboards, results, web-first, channel-neutral.

Kernel (ICNLI runtime)

Resolves your context, who you are, what you're asking, in what language, routes the message to the right extension, executes, writes back.

Extensions

Your 300+ connected apps, each a small module exposing tools the AI can call and a skeleton section describing live state. You install the ones you need.

LLM, yours or ours

The reasoning substrate. Claude, GPT, your local model, any ICNLI-compatible compute, swappable per purpose. This is where BYOLLM plugs in.

The open standard

What ICNLI actually specifies.

Open protocol at icnli.org (CC BY-SA 4.0). Where the 5-layer stack above describes how the Imperal Cloud runtime is built, ICNLI defines the 9 levels of operational context and the 5 safety laws that power the 360° AI Exoskeleton before any model touches your systems.

9 levels of context

Platform, Actor, Account, Services, Servers, Apps, Resources, Relations, Interconnections, the operational picture any AI must see before it's allowed to act.

The TWO-STEP Safety Protocol

Every destructive or high-stakes action is gated by a mandatory two-step confirmation built into the architecture, not dependent on prompt engineering, not optional.

Proactive user intelligence

The runtime notices when your context changes and surfaces it without being asked, not a chatbot waiting for the next message.

Channel neutrality

The same protocol works over Web Panel, Telegram, Terminal CLI (Webbee Code), external IDEs via Imperal MCP, voice, or any future interface. The mechanism doesn't change because the surface did.

Intent domain router

Your request gets directed to the right domain capability across 300+ extensions instead of one general-purpose agent trying to guess everything at once.

The Cloud Kernel

A distributed runtime that behaves like one machine.

Fully decentralized runtime, not one process

Decentralized kernel nodes coordinate as one unified Superintelligence over crypto-resilient encryption protocols across every contour. Any node can handle any agent or message with zero single point of failure.

Crypto-resilient contour & zero downtime

Every hop between your terminal agents (imperal-mcp), Webbee, and your infrastructure is sealed with crypto-resilient encryption protocols, while extensions and kernel upgrades roll out live without restarts.

Regional flexibility

Instances can live in different datacenters, different countries, different continents, the kernel is the logical machine across them. Regulatory and latency concerns handled at the architecture layer, not bolted on after.

Bring your own LLM

Point Webbee at Claude, GPT, an OpenAI-compatible endpoint, or your own local Ollama instance. Your reasoning substrate is swappable per tenant, live today, not a roadmap slide, and on Enterprise the per-action AI run cost drops to $0 because you're paying your own model provider directly.

FAQ

Frequently asked questions.

What is the ICNLI protocol?
ICNLI (Infrastructure Contextual Natural Language Interface) is the open specification published at icnli.org under CC BY-SA 4.0. It defines how natural language intent maps to schema-validated, consent-gated actions with cryptographic receipts.
Why does Imperal separate the Cognitive Kernel from the LLM?
Models are probabilistic reasoners; they should propose actions, not enforce security policy. The deterministic Imperal Kernel validates schemas, scrubs secrets, enforces Propose -> Confirm gates, and signs audit logs outside the model.
How does the live AST code graph work?
Imperal indexes repositories into a symbol and dependency graph (`orient`, `graph_slice`, `impact_of_change`, `read_symbol`). Agents read exact functions and downstream callers in one call instead of burning tokens scanning whole files.
Is Imperal locked to a single AI model provider?
No. The 5-layer stack treats the LLM as a swappable reasoning engine. You can route tasks across Claude, GPT, Gemini, or private open-weight models without changing a single extension or security rule.

Build on it, or just use it

The protocol is open. So is the door.

Use Webbee today with the apps already in the marketplace, or write the next extension yourself with the SDK.

Explore the SDK