Quick start

This page walks the whole journey once, end to end: from creating an account to watching an agent answer a real question about your own data. Budget about fifteen minutes. Every step has a deeper page linked at the bottom.

What you are creating

An agent is a named assistant that answers questions about a specific slice of your business data. It is not a general chatbot: it is bound to metrics you approved, follows skills you selected, and refuses questions outside the scope you defined. That is what makes its answers safe to forward to a colleague or a client.

The people who create agents are builders. The people who only ask them questions — your team, or your clients in the Portal — are members, and they cannot change how an agent works.

1. Create your account

Register at /register with an email and password, then sign in. A personal workspace is created for you automatically, so there is nothing to configure before your first upload.

A workspace owns everything you create: datasets, metrics, skills, agents, members and usage. If you belong to more than one workspace, the switcher in the top bar shows which one you are working in — always check it before uploading data.

2. Get data in

Open Data (/data) and upload a CSV, TSV, XLSX or JSON file. If an XLSX workbook has several sheets, choose the sheet to import; the whole workbook is not imported as one table. You can also provide a public HTTPS URL that returns CSV or JSON directly.

After the data lands, three things happen without you asking:

  1. The rows are loaded into the query engine, so they can be aggregated.
  2. Each column is profiled — its type, semantic role (date, category, measure or identifier), and useful statistics.
  3. The system drafts metrics, dimensions and starter views for your review.

Review the proposals and approve the metrics that match how your business actually measures things. Add clear names, descriptions and formulas. Approved metrics become governed business definitions and the vocabulary your agents use; catalogs organize them by subject area. A catalog subscription follows changes to that catalog, while individually selected metrics remain an explicit selection.

If your numbers already live somewhere else, connect Google Sheets, Feishu Bitable, S3 or Snowflake and import a snapshot. HTTP/API, email and webhook connections are tools or delivery and verification capabilities, not automatic data-source connectors. See Prepare your data.

3. Create an agent

Open Agents and start a new one (/agents/new). The wizard asks four questions:

  1. Catalog — which subject areas (sales, marketing, finance…) the agent may read from, or which individual metrics and workspace catalog it inherits.
  2. Skills — the reusable analysis playbooks it should follow.
  3. Scope — its purpose, example questions it must handle, and what it should do with questions it must refuse.
  4. Name and review — a name and description, pre-filled from your answers.

You land on the agent's build page as soon as it is created. Full detail in Create an agent.

4. Ask it something real

Open the agent's chat and ask a question you already know the answer to. That is the fastest way to catch a wrong join or a mislabelled column.

Every grounded answer carries its own audit trail: expand How I worked this out to see the metrics and data steps actually used, then open the adjacent verification entry for its sources and metric references. If an answer has no data behind it, the panel says so rather than inventing a citation. Recoverable tool failures may be retried automatically; only a failed run is called out prominently.

When an answer is wrong, report it from the message itself. The report lands in your feedback inbox with the question and the answer attached, so you can fix the metric rather than re-litigate the conversation.

5. Prove it, then ship it

Before anyone else relies on the agent, run its evals — a set of test questions generated from the scope you wrote, including questions it is supposed to refuse. Pass rate and coverage are quality signals that help catch regressions, not a standalone publishing gate. The Readiness tab separates those signals from the actual publishing requirements: a defined scope, documented metrics and reviewed refusal cases.

After the agent completes one successful run and meets those publishing requirements, publish it to your Portal and invite the people who should use it. They get a clean chat surface, no builder controls, and only the agents you published. See Publish to the Portal and Invite your clients.

Where to go next