
Nothing Died. But It All Changed.
July 31, 2026
Sigma Agents and Snowflake Agents: What They Are and When to Use Each
October 8, 2026Building with Sigma + Claude
(….and how BI development has changed, and what it means for your team)
I started my consulting career building Tableau dashboards. These were the golden days of the “data fam” when you dragged green and blue pills to x and y axes. The commercial aspect of it was pretty straightforward. I knew Tableau better than 99% of users, so companies would pay for me to build better dashboards in a more expedient manner than what they could get done internally.
Fast forward 10 years, and this type of work is quickly disappearing.
The work of building data solutions to answer business questions is not disappearing – how the work is completed and its delivery methods are changing. I’ve worked at a lot of large companies. Never once did I complete a project or a sprint and hear the words – “that’s all folks, we’ve accomplished everything in the backlog.” Timelines for building dashboards will rapidly compress and coding will be commoditized, but understanding how to build holistic solutions that drive business outcomes will still be a very important skill.
As an entire industry’s ways of working change, it’s important that our technology partners are adapting to that change. One partner in particular – Sigma – has been bobbing and weaving with these changes better than most. Once viewed as an up-and-coming competitor to Tableau and Power BI, they’ve completely shifted their approach to view themselves as an app building platform connected live to a governed data warehouse that enables users to build dashboards, apps, and Excel-like sheets all with the power of AI.
In this blog we’re going to be discussing how we leverage Sigma and Claude integration at our clients and what it means for consultant skillsets going forward.
The Sigma MCPs
At any client kickoff, access to a seat in the client’s Claude account is a checklist item. That is quickly followed by the setup of Sigma MCPs in their tenant if it’s not done already.
Sigma offers two MCP servers: the Sigma Documentation MCP and the Sigma MCP. The Documentation MCP gives Claude a searchable reference library of Sigma’s help docs and API reference, so developers get accurate guidance on formulas, features, and endpoints. The Sigma MCP lets Claude work inside your Sigma organization, as you and with your permissions, to search, explore, and query your data models and workbooks, and to chat with workbook agents. Building data models and workbooks happens separately, through Sigma’s “as code” capabilities in Claude Code, which we cover below.
| Sigma Documentation MCP | Sigma MCP | |
|---|---|---|
| What it does | Searches Sigma’s help docs and API reference | Searches, explores, and queries your Sigma org; chats with workbook agents |
| Tools | searchDocs | Search, Explore, Analyze, Agent calls |
| URL | https://help.sigmacomputing.com/mcp | Unique to your org (Profile > Integrations) |
| Authentication | None | OAuth, as the signed-in Sigma user |
| Touches your data? | No | Yes, live and governed |
The Documentation MCP
AI is only as good as what it knows. Early AI models without MCP context would regularly invent new functions or inappropriately apply functions that did exist. The Documentation MCP ensures that Claude references Sigma’s documentation before it recommends solutions for Sigma.
The Documentation MCP needs no login and never touches customer data, so we add it to every developer’s setup on day one.
The Sigma MCP
The Sigma MCP lets Claude search, explore, and query your Sigma organization as the signed-in user. That means anyone can ask a business question in plain English in a Claude chat window and get an answer built on the same governed definitions as your dashboards.
It works best pointed at data models, where metrics and relationships are already defined. It only queries published content, and a few advanced calculations are not supported yet.
How authentication works between Sigma and Claude
Claude borrows your Sigma identity through OAuth, and Sigma enforces everything from there on.
- A Claude Owner on a Team or Enterprise plan adds the Sigma connector for the whole organization, or an individual adds it from the Connectors Directory.
- The user selects Connect. Claude sends them to their Sigma org to sign in with whatever login they already use.
- The user authorizes Claude, and Sigma issues a token tied to that user.
- Every search and query runs as that user, with their account type, their document access, and their row-level security.
If Sigma has an IP allowlist turned on, add the IP addresses Claude uses for outbound connections with the scope set to API.
Setting up Sigma MCPs
In Sigma (an admin, once):
- Enable Use Sigma MCP with OAuth on the account types that should have it.
- Confirm those users have View connections, View data models, and/or View workbooks, plus at least Can view on the documents and Can use on the connections they will query.
In Claude (each user):
- In Sigma, open your user menu > Profile > Integrations, and under Connect Sigma to AI tools copy your Sigma MCP URL.
- In Claude, find the Sigma connector in the Connectors Directory and select Connect. Paste the URL if prompted.
- Sign in to Sigma and authorize Claude.
Developers working in Claude Code can add the same two servers from the terminal. Sigma’s setup guide walks through it.
Data Models as Code
Data models as code is generally available. Every Sigma data model can be pulled down, edited, and pushed back as a JSON or YAML file through Sigma’s REST API, which means you can now use tools like Claude Code to develop, iterate, and publish your data models directly to a Sigma workspace. That file holds everything the model defines: its source tables, calculated columns, metrics, relationships, and descriptions.
Once a data model lives in a file, it can live in your team’s Git repository (GitHub, GitLab, Azure DevOps, and self-hosted servers all work), and that brings branches, pull requests, and code review to your semantic layer. Developers work in their own branches instead of directly in production, every change has an owner and a reason, and a bad change can be rolled back instead of rebuilt from memory.
The catch is that the speed comes with a new way of working. Your team is now operating like software engineers, and for a lot of people coming from legacy BI, that takes training and a real change in mindset.
How we develop data models with Claude
In practice, we treat Claude Code like a developer who knows Sigma’s spec format cold. We describe the change, Claude writes it, and a person reviews it before anything ships.
The setup is light. Sigma’s plugin for Claude Code (released May 2026) teaches Claude the data model spec and handles sign-in, so each developer authenticates once with their own Sigma login and works from a project folder on their laptop that is connected to the client’s Git repository. Automated deploys run under a dedicated credential with a named owner, and nobody hands Claude a production admin key.
From there, a typical change looks like this:
- Pull. “Pull down the Revenue data model.” Claude grabs the latest version from Sigma.
- Describe the change in plain English. “Add a Gross Margin % metric, connect Orders to Customers, and write a description for every column that is missing one.”
- Claude builds and checks. It writes the change, checks each formula against the Documentation MCP, and summarizes exactly what changed.
- Preview. Claude publishes to a dev folder in Sigma so we can click through it the way a user would.
- Review and ship. A teammate approves the pull request, and the change deploys to production.
Where this saves the most time:
- Bulk cleanup. Standardizing column names, descriptions, and formatting across dozens of data models in one pass.
- New models. Standing up a data model from warehouse tables, with relationships and metrics included, in a fraction of the time.
- Migrations. Translating legacy BI logic, like Tableau calculations or Power BI measures, into Sigma formulas.

Building Workbooks as Code
Workbooks as Code, now in public beta as of September 2026, applies the same approach to the dashboards themselves. We write a brief, and Claude builds the workbook on top of a governed data model.
This is the part that hits closest to home for me. Most of that old Tableau build was translation: requirements into calculations, calculations into charts, charts into a layout. Translation is exactly what Claude is good at, as long as it is working from a governed data model rather than raw tables.
How we do it:
- Data model first. Every workbook sits on a governed data model, so metrics are defined once and the workbook only arranges them.
- Write the brief. Plain English: who uses it, what decisions it supports, which metrics matter, and how people need to filter it.
- Prompt for the first draft. “Using only metrics from the Sales data model, build a two-page workbook from this brief, and tell me anything you could not build.”
- Iterate in plain English. “Move the trend chart above the table.” “Add a region control.” Each round takes minutes, not days.
- Review and ship. The workbook goes through the same pull request flow as the data model.
What to expect while it is in beta:
- Specs are big. One big workbook can easily be well over 10,000 lines of code. That is a job for Claude, not a person with a text editor.
- Updates replace the whole workbook. Always start from the latest version.
- Not everything is supported yet. Validation is stricter than the UI, and some chart types still need to be built by hand.
Existing Sigma customers can ask their account team about early access.
How security fits together
Every path into Sigma authenticates as a specific identity, every request passes through Sigma’s permissions, and queries run live in your warehouse. No extracts or copies are created, and Claude only sees results the user is already allowed to see

Business users ask questions in Claude, and the Sigma MCP answers as that user, ideally by querying a governed data model. Developers build specs locally in Claude Code, push them to your team’s Git repository for review, and only an approved pull request changes production.
What this means for clients (and for consultants like us)
The dashboard that took me 2 to 4 weeks is becoming a few days of brief writing, prompting, and review. The backlog does not shrink. The bottleneck moves.
The hours now go to the work AI cannot do on its own: designing a semantic layer people trust, writing briefs that capture the real business question, reviewing what Claude built with a critical eye, and owning the governance that keeps it all safe. Knowing where the green and blue pills go is no longer the differentiator. The differentiator now is designing workflows, with tools like Sigma and Claude, that do more than answer business questions. They drive action toward business goals.
Tools like Sigma also change what we are building in the first place. A dashboard tells you what happened. Sigma workbooks can now take the next step themselves. Input tables write decisions back to the warehouse, workbook actions can call stored procedures or external APIs to kick off a process in another system, and Sigma agents can run those actions on a user’s behalf, with an approval step when you want one.
Our job is shifting from building places to watch KPIs to building workflows, agents, and system connections where taking action is the default way of working, not just observing the data.
Instead of downloading a desktop BI tool, these are the actions we are taking or discussing during client kickoffs:
- Connect the MCPs. The Sigma Documentation MCP for every developer, and the Sigma MCP for business users once account types and row-level security have been reviewed.
- Move data models into Git. A client-owned Git repository, pull request review on every change, and automated deploys to production.
- Pilot Workbooks as Code. One or two high-value workbooks, wherever beta access is available, built from a brief and promoted through the same process.
Contact Us
If you are trying to figure out where Sigma, Claude, and your data team fit together, that is exactly the conversation we like having. Please email us at spencer@maverickdata.io!



