Governed AI · Observed AI

AI that operates in front of real customers, never left unattended.

Magnus is the foundation for building applications and agents with AI at the center. The model understands and writes; decisions follow rules, anything irreversible gets confirmed, and every step is recorded.

Explore the products Let's talk

Inside a turn restaurant

Table for 4 on Friday at 9pm?

Understands Model

party: 4 · day: Friday · time: 21:00
Decides Rules

"booking" flow → check availability
Executes Tool

availability(fri 21:00) → table free
Confirms Gate

booking is irreversible: asks for a yes first

I have a table for 4 on Friday at 9pm. Shall I book it under your name?

Trace recorded: what it understood, which rule applied, what it ran.

The problem

AI already talks to your customers. Who is supervising it?

Getting a language model to chat is easy. Letting it operate unsupervised, in front of an end customer, is something else.

It makes things up

It states a policy, a price or an opening time that does not exist.

It acts without permission

It confirms a booking, an order or a payment nobody approved.

It leaves no trail

When something goes wrong, there is no way to reconstruct why it did what it did.

It fails silently

A provider goes down or a conversation gets stuck, and nobody notices until the customer complains.

Your options

Today a business has two paths. Magnus is the third.

Three questions are enough to compare them.

Path 1

An AI chatbot

Fast, but nobody trusts it with anything that matters.

How long until it works?: Days
Can you trust it to operate?: No: it makes things up and acts unchecked
Does it account for what it did?: No record of why

Path 2

Custom software

Reliable, but slow and expensive.

How long until it works?: Months, with a dedicated team
Can you trust it to operate?: Yes, it does exactly what was specified
Does it account for what it did?: Only if built for it

Path 3

Magnus

AI that operates by the rules, without starting from scratch.

How long until it works?: Weeks: you configure, not code
Can you trust it to operate?: Yes: rules, and a confirmation before anything irreversible
Does it account for what it did?: Yes: every step is recorded

Already using OpenAI?

Change the address. Magnus takes control of your AI.

If your application already talks to OpenAI, nothing gets rewritten: the same code points at Magnus, with your Magnus key. The model still understands and writes; what to do is now up to Magnus.

Before api.openai.com/v1

Now app.iamagnus.com/v1

Magnus keeps the conversation: it reads only the new message, and user keeps each customer’s apart. Tools are configured in the agent, not per request, so tools and response_format are not accepted.

Get an API key

 from openai import OpenAI  client = OpenAI(-    api_key="sk-...",+    base_url="https://app.iamagnus.com/v1",+    api_key="magnus_sys_...", )  reply = client.chat.completions.create(     model="gpt-4o",     messages=[{"role": "user", "content": "A table on Friday?"}],+    user="jane@company.com",  # one conversation per customer ) print(reply.choices[0].message.content)

+3 −1 lines changed · Python

 import OpenAI from "openai";  const client = new OpenAI({-  apiKey: "sk-...",+  baseURL: "https://app.iamagnus.com/v1",+  apiKey: "magnus_sys_...", });  const reply = await client.chat.completions.create({   model: "gpt-4o",   messages: [{ role: "user", content: "A table on Friday?" }],+  user: "jane@company.com", // one conversation per customer }); console.log(reply.choices[0].message.content);

+3 −1 lines changed · Node.js

Our approach

The model understands. The system decides.

We use the language model where it shines: interpreting what a person says and writing the reply. What to do is decided by a deterministic engine, with declared rules anyone can read and audit.

Governed

Explicit rules per industry. Nothing irreversible without a confirmation. Handoff to a person when it is called for.

Observed

Every turn leaves a trace: what it understood, what it decided and by which rule, which tool it ran and what came back.

Resilient

If an AI provider fails, it moves to the next one. Under pressure, the system degrades in a controlled way instead of breaking.

Products

Three levels. Start with the one you need.

Each level includes everything in the one before. Choose by how much you want to build yourself and how much you want solved for you.

Level 1

Magnus Zero

The engine that decides

For: Developers

Rule chain and state machine
Confirmation before anything irreversible
Your own tools and MCP
Failover across AI providers
Runs with no database and no API key

Open source · coming soon

Level 2

Magnus Core

The engine, production-ready

For: Engineering teams and integrators

Everything in Zero, plus:

OpenAI-compatible API
Organizations with isolated data
Users, roles and API keys
Knowledge base
Per-turn traces
SDKs for Python, Node and Go

core.iamagnus.com

Level 3

Magnus Agents

Agents ready for your business

For: Businesses, no code needed

Everything in Core, plus:

Pre-configured agents per industry
WhatsApp, Telegram and web
Live conversations
A person steps in when needed
Dashboard and metrics

agents.iamagnus.com

Agents is what you see.Core is what thinks.Zero is how it decides.

Integration

Magnus plugs into your product. It does not replace it.

Your application stays yours, with your interface and your brand. Magnus works behind it, between your application and your systems.

1
Your application

Sends your customer’s message.
2
Magnus Core

Understands, decides by the rules, confirms anything irreversible and records every step.
3
Your systems

Your calendar, stock or CRM: the agent queries and operates them as tools, through MCP or its own.

The answer goes back to your application, ready to show.

Starting from scratch? A few lines, in your language

pip install iamagnus GitHub →
npm install iamagnus GitHub →
go get github.com/ABZ-LABS/magnus-go-sdk GitHub →
Request an API key

Python

from iamagnus import MagnusClient

with MagnusClient("https://app.iamagnus.com", "magnus_sys_...") as client:
    agent = client.list_agents()[0]["id"]

    # One thread per end user: `user` continues it, not resent history.
    chat = client.conversation(agent, user="jane@company.com")

    print(chat.send("Hi, what can you do?"))
    print(chat.send("And the price?"))

Node / TypeScript

import { MagnusClient } from "iamagnus";

const client = new MagnusClient({
  baseUrl: "https://app.iamagnus.com",
  apiKey: "magnus_sys_...",
});

const [agent] = await client.listAgents();

// One thread per end user: `user` continues it, not resent history.
const chat = client.conversation(agent.id, { user: "jane@company.com" });

console.log(await chat.send("Hi, what can you do?"));
console.log(await chat.send("And the price?"));

client := magnus.New("https://app.iamagnus.com", "magnus_sys_...")
ctx := context.Background()

agents, err := client.ListAgentsContext(ctx)
if err != nil {
    log.Fatal(err)
}

// One thread per end user: `user` continues it, not resent history.
chat := client.Conversation(agents[0].ID, "jane@company.com")

answer, _ := chat.SendContext(ctx, "Hi, what can you do?", nil)
fmt.Println(answer)

The market

Where we stand

Each available approach solves part of the problem. Magnus targets the empty quadrant: quick to launch and safe to operate.

high Control when operating with customers → low

Custom software Full control, months of development

Agent frameworks Flexible: each team builds its own control

Autonomous agents Powerful, unpredictable in front of a customer

Chatbot builders Live in hours, no guarantees to operate

Magnus Configured in weeks, operates by the rules

Speed to production → high

Custom software: Full control, months of development
Agent frameworks: Flexible: each team builds its own control
Autonomous agents: Powerful, unpredictable in front of a customer
Chatbot builders: Live in hours, no guarantees to operate
Magnus: Configured in weeks, operates by the rules

Illustrative positions by category, not a measurement.

Standards

Moving with the industry

We do not invent our own protocols where standards already exist. We adopt them and add governance on top.

Core

OpenAI-compatible API

Anything that already speaks that chat format speaks to Magnus, without rewriting the integration.

Zero

Model Context Protocol

External tools both ways: Magnus uses them and can also be offered as a tool.

Zero

Declared flows

The industry is moving from agents that improvise to stateful flows. In Magnus a flow is a file you can read.

Zero

Human in the loop

Confirmation and handoff are part of how a turn is designed, not a later patch.

Zero

Multi-provider

No lock-in to one model: if a provider fails, the turn continues with the next.

Core

Per-turn traces

Any conversation can be rebuilt step by step: what was understood, what was decided and why.

Zero

Reproducible evaluation

Scripted dialogues and automated checks that catch a regression before it reaches a customer.

Zero

Open source

The engine is released under Apache-2.0, so anyone can verify how it decides without taking our word for it.

Use cases

Already built and running

business domains configured

official SDKs: Python, Node and Go

93.3%^*

task completion in our benchmark, against 60.0% and 65.8% for two conventional approaches

* Internal benchmark: 120 dialogues across 7 domains. It was designed by the same team that built Magnus, and the baseline systems run without tuning to the dataset. An independent evaluation on unseen data is still pending.

Built on the platform

AI concierge

Answers the building intercom through its phone system, talks to whoever is at the door and transfers the call to the right apartment. It does not decide to open the door: it decides who to route to, and logs it.

Building management

A management system for the administration and an agent for residents, working on the same data.

Real estate

Property search, visit scheduling and lead capture, using the listings and time slots the system returns.

Orders: restaurants and pharmacies

Step-by-step order taking, explicit confirmation before closing it, and handoff when a sensitive question comes up.

Customer support

Support that detects urgency and frustration, and hands off to a person with the conversation context.

Personal finance

In development

Financial projections for banks and credit unions, with each user’s data kept isolated.

Contact

Let's talk

Incubators and investors

A platform that is built and running, with a clear thesis on how to bring AI into real operations.

Request the executive brief

Partners and integrators

Add governed agents to your projects with Core, the SDKs and white-label for each client.

Propose a partnership

Businesses

Tell us what you want to automate and we will show you what it looks like running on Magnus.

Book a demo

AI that operates in front of real customers, never left unattended.

AI already talks to your customers. Who is supervising it?

It makes things up

It acts without permission

It leaves no trail

It fails silently

Today a business has two paths. Magnus is the third.

An AI chatbot

Custom software

Magnus

Change the address. Magnus takes control of your AI.

The model understands. The system decides.

Governed

Observed

Resilient

Three levels. Start with the one you need.

Magnus Zero

Magnus Core

Magnus Agents

Magnus plugs into your product. It does not replace it.

Starting from scratch? A few lines, in your language

Where we stand

Moving with the industry

OpenAI-compatible API

Model Context Protocol

Declared flows

Human in the loop

Multi-provider

Per-turn traces

Reproducible evaluation

Open source

Already built and running

Built on the platform

🚪AI concierge

🏢Building management

🏠Real estate

🛒Orders: restaurants and pharmacies

🎧Customer support

📈Personal finance

Let's talk

Incubators and investors

Partners and integrators

Businesses

AI concierge

Building management

Real estate

Orders: restaurants and pharmacies

Customer support

Personal finance