Back to homeMGA AgentHub

See how an AI agent is built,
and try it live.

We walk you through how we configured the MGA Assistant, the same one that answers on this site, with its knowledge, its calculation tool and its database. After each step you can try what was configured.

Create a trial account

Proof of concept: do not enter real personal data. Sign-up is limited to an allow-list.

Builder and chatlive
  • Identity
  • Knowledge · 29 articles
  • Tools · 1
  • Data · 18 tables
  • Publish

What is my net pay on 4,000,000 gross?

router → toolcalcular_ganancias

Estimated monthly net: $3,280,178 (Jan-Jun 2026 period).

The path of every question

  1. Question
  2. Rules
  3. Router
  4. Knowledge · Tool · Data
  5. Answer
The story in 8 steps

From a blank page to a published agent

  1. Identity

    First we decide who it is

    An agent starts with a name, a role and a voice. We took the rules of the assistant we already had in n8n and moved them into the agent instructions, so they no longer depend on the browser.

    Name
    MGA Assistant
    Role
    Technical and commercial assistant of MGA Tech Consulting
    Tone
    Professional, technical, clear and friendly
    Language
    The one the person writes in
    Ask the assistant:What services does MGA offer?
  2. Rules

    Then we set its limits

    A public assistant has to know what not to do. We defined the forbidden topics, a few rules that are never broken and the message it uses to decline politely.

    Forbidden topics
    Weather, politics, history, jokes, programming unrelated to the services
    Strict rules
    Never invent information · Never name the team or its size
    If it goes off-topic
    Declines with its own message and goes back to the services
    Ask the assistant:How is the weather in Buenos Aires today?
  3. Knowledge

    We give it something to read

    We loaded the 29 technical blog articles, the same ones the n8n agent consulted. The platform splits them into fragments and looks up the closest ones before answering.

    Source
    29 blog articles (Markdown)
    Result
    233 fragments ready to query
    Search
    Lexical and local, no external services
    Ask the assistant:What do you recommend to reduce BigQuery costs?
  4. Tools

    We give it something to do

    Income tax math cannot be made up by a model. We turned it into a tool the agent calls with the data the person gave, and that is read-only.

    MCP server
    mga (internal)
    Tool
    calcular_ganancias · read-only
    Who computes
    The tool, with the January-June 2026 brackets
    { "sueldo_bruto_mensual": 4000000,
  "hijos": 1, "conyuge": true }
→ neto_mensual_estimado: 3280178
    Ask the assistant:I earn 4,000,000 gross per month, with a child and a spouse. What is my net pay?
  5. Data

    We connect it to a real database

    So it can answer with portfolio and pack data, we connected it to PostgreSQL with a read-only user. The model proposes the query and the platform validates and limits it before running it.

    Database
    PostgreSQL, read-only user
    Allowed tables
    18 of 59: portfolio, packs, blog and taxes
    Left out
    Contacts, students, payments and conversations
    Limits
    SELECT only, with time and row caps
    Ask the assistant:What consulting packs do you offer?
  6. Look

    We give it the brand look

    We picked the MGA colors, a short greeting and four quick replies to get started. We placed it at the bottom left so it lives next to the current assistant.

    Colors
    Cyan #22d3ee on dark #070b12
    Position
    Bottom left
    • Calculate my income tax
    • What services do you offer?
    • I want to automate a process
    • Contact the team
    Ask the assistant:I want to automate a process
  7. Publish

    We publish it with one line

    We authorized the domains where it may answer and copied a one-line script. From any other site, the agent does not answer.

    Domains
    www.mgatc.com, mgatc.com and the previews
    Widget size
    11.6 KB, isolated in Shadow DOM
    <script src="https://agents.mgatc.com/widget/agent.js"
        data-agent="<ID-DEL-AGENTE>" async></script>
    Ask the assistant:Hi, what can you do for me?
  8. Test and measure

    We test, look and tune

    Every conversation leaves step-by-step traces: what the router decided, which fragments it read, which tool it called. With that we tune the agent and control spend before every model call.

    Traces
    Router, knowledge, tool and data
    Analytics
    Conversations, resolution and most asked topics
    Costs
    Credits and spend cap, with optional own key
    Contacts
    Handed to a person by email
    Ask the assistant:I want to talk to someone from the team

You saw how it was built. Now try it.

Ask it to compute your income tax, search the blog or tell you which packs we offer.

From the n8n flow to platform components

In n8nIn MGA AgentHub
System prompt built by the siteAgent identity, tone and instructions
Off-topic refusal inside the promptRules: forbidden topics and refusal message
Keyword intent detectionIntent router with a low-cost model
RAG with embeddings and Supabase (29 articles)Knowledge: the same 29 blog articles
Income tax math in a code nodeMCP server with the calcular_ganancias tool
Supabase reads with nodesRead-only database connector
Lead saved in Supabase plus alert emailContacts and email hand-over
Fixed model on Hugging FaceModel set per platform or with your own key