Skip to main content
Preview You want people to ask a data question in any AI tool and get the number your team agrees on. Every piece of that has its own page on this site, and together they cover more than one pilot needs. This page puts the pieces in the order teams do them, says what you should see after each one, and lists where people get stuck. It doesn’t repeat the steps. Each stage links to the page that has them.
At the end of this lesson: your team has released metrics in a shared workspace, your agent reads them before it writes SQL, and two eval runs on your golden questions show what changed.

Two ways to start

The Analytics plugin has two starting points. Pick one before you plan the week. The demo ends with Continue building on real data, so you can do both. See Try demo. If you already have an analytics agent and a set of questions with known answers, your week is different. You don’t rebuild the agent in Atlan Trident. You score it as it is today, connect it to your metrics, and score it again on the same questions. Do stage 6 once before stage 5 and once after.

Before you begin

Before you start, make sure you have:

Team

  • A small group. Two or three people who will build, and one person who knows what the numbers should be. One person alone can build metrics but has nobody to settle a disputed definition with.
  • One use case. A business area and the questions people ask about it, such as “online store orders: revenue, order count, average order value”. A specific request gives better metrics than “all of sales”.
  • A coding agent on each builder’s computer. Claude Code, Codex, or Cursor, with Python 3. See Set up Analytics in your coding agent for the full list.
  • A warehouse with real query history. The richest source is the SQL your team already runs. A development warehouse that nobody queries has none. If that is all you can reach, start from a Power BI or Tableau file, or from a business question.
  • Golden questions, if you have them. 15 to 30 real questions, each with an answer an analyst has confirmed. If you have none, write them in stage 6.

For whoever grants access

For IT and security

  • The web app needs no install. People open app.atlan.com in a browser. Metric Studio, evals results and traces all work there.
  • Atlan Trident Desktop is optional for this path. It adds two shortcuts: signing in installs the Atlan plugin into the AI tools on that computer, and Get started in Metric Studio opens the build prompt in Ask AI for you. If your security software blocks the installer, use the web app. See Web app and Atlan Trident Desktop.
  • The plugin runs on the builder’s computer. It reads your warehouse and BI files with the access that person already has. Metric Studio never connects to your warehouse and never runs SQL. What reaches Atlan Trident is what the builder approves: metric definitions, saved queries, and the links between them. Prepare data access has the detail a security review asks for.
  • Coding agents reach Atlan Trident at one address, https://api.atlan.com/mcp, and each person signs in through their browser. See Network requirements if you restrict outbound traffic.

Plan your week

1

Set up access

The first admin signs in, creates one shared workspace for the pilot, invites the group into that workspace as Builder, and installs Metric Studio. Decide now whether metrics an agent saves should reach other agents straight away: a workspace publishes automatically unless you set Publishing to Manual.You should now see: each builder opens Metric Studio, chooses Metrics, picks the shared workspace, and sees Build your first metrics. If you can’t start a build there, you can read the workspace but not create in it: see You can’t create or edit anything.
2

Connect coding agent

Each builder installs the Analytics plugin, atlanai-analytics, in the coding agent they already use, then reloads the agent. It works with the atlanai plugin from the same marketplace, which holds the connection to Atlan Trident.Sign-in can wait. You need it only when you save metrics to Atlan Trident or ask questions of released ones.You should now see: in Claude Code, /mcp lists a server named plugin:atlanai:atlan. After you sign in, ask “List my Atlan workspaces.” and the shared workspace is in the answer.
3

Build metrics from real queries

In a project folder, ask the agent to set up analytics on your data, and name the business area. The plugin shows its steps and keeps them up to date. It surveys your sources, asks what you want to answer, builds metrics and saved queries, and runs each one on your data. Nothing is written to Atlan Trident during this stage.When it asks where the context should live, choose the shared workspace, not your personal one. The plugin recommends your personal workspace, which only you can read.You should now see: a “CONTEXT BUILT” summary in the chat with the number of metrics and queries, and how many were checked on your data. Before anything is saved, the plugin shows the plan with its counts and waits for your approval.
4

Review and release

Open Metrics in Metric Studio. Read each metric’s Overview against what you know about the number. Ratios, joins and grain are the parts most likely to need a correction. Then release the ones you trust.Under Automatic publishing, the default, metrics your agent saved are already released when you get here. Review them anyway, and withdraw or fix any that are wrong.You should now see: in the Metrics list, a released metric shows Live and a version such as 0.1.0. The rest show Not released.
5

Point your agent at Atlan Trident

Agents read metrics through MCP Gateway. You don’t rebuild your agent: you add one connection, and the agent searches for a metric and reads the released version before it writes SQL. It runs the SQL itself, with its own warehouse access.Pick the page for where your agent runs:You should now see: ask a question a released metric answers. The agent’s registry__get_artifact result has a release.version_label that matches the Live label in Metric Studio.
6

Score it on golden questions

Store your questions as a dataset, wrap your agent as a function that takes a question and returns its answer, and score each case twice: is the number right, and did the agent read the metric it should have. The first run is your baseline.If you already have an agent, run this once before stage 5 with a candidate name such as before-metrics, and again after. The comparison shows what the metrics changed.You should now see: the script prints an experiment ID and a mean for each score. In the app, open Evals, then Experiments, and select the experiment to see one case per question.
7

Read traces and fix

For each failed question, the two scores say where to look. A wrong number from the right metric points at the formula, filters or grain. A missing metric read usually means the question uses words the metric doesn’t. Open the case’s trace to see what the agent searched for, which metric version it read, and the SQL it ran.Fix the metric in Metric Studio. Saving creates a new revision, and agents keep reading the released one until you release again. Run the same questions with a new candidate name before you release to everyone.You should now see: two experiments on the same dataset, and the questions whose scores changed between them.

Where people get stuck

What you have accomplished

You set up a shared workspace, built metrics from your team’s real queries, released the ones you trust, connected your agent to them, and scored it on golden questions before and after a fix.

What’s next

Evaluate your analytics agent: Go deeper on scoring, reading failures, and checking a fix.

See also