George Salmon
Technical Consultant at Bulien
Roan Pilsworth
Technical Consultant at Bulien
If your team uses Alteryx One, someone has probably asked whether you can just point Claude at it. The answer is yes, and the connection is the easy part. The harder questions are what the model should be allowed to see, and what you want it to do once it can see it.
Our view is that the most useful thing an external model can do with the Alteryx MCP server is run workflows you have already built and trust. Raw datasets are an option, but a short list of checked processes is a much better place to start.
MCP stands for Model Context Protocol. In our video it is described as a cookbook of recipes that tells a language model which steps it is allowed to take for a given task. Connected to Alteryx One, it lets the model look at datasets and run workflows in your own environment.
It sits inside Agent Studio, a new section in the left hand navigation. At the time of recording (the video was published in August 2026) Agent Studio had three tabs: Alteryx agents (in preview), marketplace apps (prepackaged MCP connections) and custom MCP connections, which link your own external model such as Claude Code or Codex. That last tab is the one that matters here.
Before any model can do anything, you open Set up assets and toggle on the datasets and workflows you want available. Each one you switch on becomes an endpoint that the MCP server, or an Alteryx agent, can use. Nothing else is visible to it.
That screen is your control. We would treat it like permissions on a shared drive: start with a handful of workflows that already run reliably, and add more once you are happy with how the model behaves.
Roan Pilsworth described the setup as “incredibly straightforward”, and on screen it was. In Custom MCP connections you choose the model, in his case Claude Code, and expand the dropdown for its setup instructions. These include a line you can copy into your terminal and run. Roan went one step simpler and handed the line straight to Claude Code, which loaded the MCP connection itself. An OAuth window then opened in the browser, he signed in with his usual username and password, and the connection was live. George Salmon did the same with Codex and let it connect itself. The sign-in is what gives the model access to your environment, so it matters whose account is used.
In our demonstration, Roan asked Claude to find the workflow that returns total sales value and run it. He did not name the workflow. Claude listed the workflows it had been given, worked out which one fitted, ran it and returned the answer in natural language.
Roan’s reason for running a workflow instead of querying the dataset directly is the one we would give: it is “the visible and repeatable, auditable process that we’ve already got created”. The logic is in a workflow your team can open and inspect, rather than in whatever the model decided to calculate that day.
The run behaved exactly as if Roan had pressed run himself. It used external connections through DCM, and the email tool at the end sent its email. With cloud execution for desktop, Claude could also reach desktop workflows living in Alteryx One. That cuts both ways: an email tool in an exposed workflow sends real emails when the model runs it.
The same run produced a small, confident error: Claude reported the total in dollars when Roan’s data was in pounds. In fairness, the currency was not stated anywhere in the data. Mistakes like that slip into slides unnoticed, so label your units and read what comes back.
Checking matters more as you layer reasoning on top. George had Codex run a workflow, then summarise the output by department and month. Useful, but that summary is a calculation the workflow never did, so treat it as new analysis.
Alteryx agents are separate from the MCP connection. George set one up on a workforce cost analysis dataset and asked it to find key insights. Each step came with a small JSON snippet showing the operations performed and how the data was filtered, which is a better audit trail than many chat tools offer.
Custom instructions changed the tone: a playful, sarcastic brief produced the headline “The IT department is having a party and you’re paying.” Chart instructions did not land, and they could not stop it producing a donut chart. At the time of recording the agents worked on datasets only and read as a repackaged Auto Insights.
In a separate video, George had Codex use the Alteryx skills files to build a containerised workflow in six sections and a six page Bulien Pulse dashboard, from over 10 HR datasets and a deliberately vague brief, using native Alteryx tools only. Chris McEleavey, who heads our technical team, works to an 80/20 rule here: AI gets you about 80% of the way, and a human in the loop gets it over the line.
George and Roan walk through Agent Studio, connect Claude Code to Alteryx One and have it pick and run a workflow from a single prompt.
We help teams decide which Alteryx workflows are safe to hand to a model, and put the checks around them so the answers can be trusted. If you are weighing up AI on top of Alteryx, get in touch and tell us which processes you have in mind.