AI setup · MCP
Recommended
Set up using AI
Connect Nekt’s MCP to your preferred AI and move forward with context, approval and security at every step.
ChatGPT
Claude
Gemini
Cursor
+ other
Self-guided setup
Getting started with Nekt
Connect a source, validate the first load, and ask your data a question in Claude, ChatGPT, or another MCP-compatible client. Set aside about 15 - 25 minutes of active work; the first load can continue running in the background.
Before this guide
This roadmap starts with a ready-made Nekt workspace. If you haven't created your account or workspace yet, do so first. workspace configuration. You can stop after any step: your progress on this page is saved in this browser.
Haven’t created your workspace yet? Watch tutorial
Need help?
Inside Nekt, click Help, in the upper right corner. The assistant consults our knowledge base to guide you through the entire process.
Open Nekt ↗
You stopped at step 1.
This is the first question of every deployment, because the answer defines how your catalog is organized. Choose below: step 2 adapts to your choice.
Why decide this now
It's the only decision on this page that takes work to undo. Change the structure later means redoing table paths, permissions, and the context the agent uses to respond. Everything else is reversible in a few clicks.
Check the items above to unlock the next step
With the structure decided, create the layers in the Catalog. Doing this beforehand prevents the first fountain falls in a temporary place that you will want to change later. Start simple: the organization can evolve with use.
bronze how it arrives from the source, without treatment
silver validated, standardized, no duplicates
gold ready for consumption and for the agent
client_a isolated, with own access
client_b isolated, with own access
client_a isolated data from customer A
client_b isolated data from customer B
internal_bronze internal data as it arrives from the source
internal_silver standardized internal data
internal_gold ready-to-consume internal data
The impact
The catalog is what the agent sees. Layer and table names are context: a question
about revenue is answered much better using a taxonomy like gold.receita_mensal than in
raw.tbl_001.
Configuring a source follows six steps: choosing the connector, credentials, selecting streams, mapping for tables in the Catalog, trigger and description. The most relevant points in the configuration are the ones below, the rest can keep the default options.
Credentials
Each source has its own format (token, OAuth, permissions, etc). Follow the connector step by step. When in doubt, consult the guide for connecting sources before testing new credentials.
Streams and tables
First select only the datasets you will use now and map each stream to a Catalog table and layer. Choose names that explain the content, such as
crm_negocios or financeiro_recebimentos, rather than hard-to-recognize acronyms.
Trigger
When you save the source, the first run starts automatically. Choose a trigger that matches how often the data changes: daily is a good starting point for data that changes once a day; use a higher frequency only when the decision really requires it.
If the credential is not with you
It is the most common reason for stopping. Ask the person responsible for access with the permissions required by the connector and resume this step when it is available. Do not use a shared personal password as a shortcut.
The first run starts automatically after the source is saved and can take from a few minutes to a few hours, depending on the extracted volume. Subsequent runs follow the sync type configured for the connector (for example, full, incremental or log-based). You can close the tab while it runs. To track progress or inspect an error, open Runs on the source details page.
Don't rely on checking manually
Configure failure alerts to the source. You can activate them yourself or, if you are owner/admin, define global recipients in Settings → Alerts. This way, a future failure does not depend on someone opening Nekt to be noticed.
With the data in the Catalog, enable Nekt's MCP Server and connect your agent. This is where you start interacting with data in natural language. MCP Server is available on paid plans.
Nekt offers guides for Claude, ChatGPT, and Cursor, as well as other MCP-compatible clients. The authentication method changes depending on the client: apps like Claude Desktop and Cursor use OAuth; Command line and automation tools can use token. In Integrations → MCP Server, enable the server and follow the guide for your AI client.
Ask the right question first
Start with the most obvious question about your business, the one whose answer you you already know it off the top of your head. Giving an answer that you can validate is what builds trust for the next ones.
These are short texts that explain your business to the agent. Create the first one with the person who knows the business metrics and rules; AI can help with drafting, but it should not decide on its own which definition is correct. It's the most underestimated step — and the one that changes the quality of the response the most.
Creating a context document is like explaining your business to a new colleague. He doesn't need to know everything, but he needs to know enough to answer correctly.
- What each metric means for your company
- Which tables and sources to use for each question type
- Rules that only insiders know
The impact
Without context, the agent responds with what it can deduce from the table and column names. With context, it responds with the definition your company uses. Two companies with the same different schema and definitions of "active customer" require different answers, and only the context document tells him this. To delve deeper, see the complete guide to context documents.
At the end of the six steps
Your team asks, and the agent responds with the company's data.
Complete the steps above and this area shows you what to do next.
What about modeling and transformations?
The goal of this guide is to get to the first reliable answer, not to model everything at once. After validating a question, use recurring questions to prioritize Queries, Notebooks and semantic layer. When saving a Query in Nekt, you define the layer and name; execution generates and updates the output table in the Catalog — there is no need to create a view manually. Start with the data or metric that the team asks about most.
If you get stuck at any step
Before calling support, isolate the problem point. This reduces back and forth and allows you to move forward on your own when it's simple.
01Source does not connect
Review the permissions and credential format in the connector guide. If necessary, request specific access from the person responsible for the source.
02Execution failed
Open the run with source error and read the log. First confirm that the credential is still valid and that the chosen stream exists.
03Agent does not access data
Confirm that MCP Server is enabled and that OAuth or token is allowed for the tables used in the question.
04Still need help
Click Help, in the top right corner of Nekt, and talk to our assistant. He consults Nekt's knowledge base to answer questions during the process. To speed up guidance, enter the step number, resource name, error message and run link, if any. Never send passwords, tokens or other credentials via chat.